LaunchDarkly vs GrowthBook

Agent access

1. ai-native userPoint an agent at llms.txt or agent-oriented docs

   weight 2 · round to GrowthBook
   LaunchDarkly—none0/10

   Direct probes show llms.txt and docs.md both return 404, and no evidence of any agent-oriented documentation endpoint; while LaunchDarkly ships an MCP server and CLI, these do not satisfy the specific 'llms.txt or agent-oriented docs' story.

   • [probe] “PROBE llms.txt: HTTP 404 at https://launchdarkly.com/llms.txt”
   • [probe] “PROBE docs-md: HTTP 404 at https://launchdarkly.com/docs.md”

   proof ↗⚑ flag

   GrowthBook✓fullprobed9/10

   A probe confirms llms.txt is live at docs.growthbook.io/llms.txt returning HTTP 200 with a structured index of 180 documentation pages, exactly matching the agent-oriented docs pattern; GrowthBook also documents an official MCP server for agent integration, reinforcing agent-friendliness. Missing for 10: no independent third-party confirmation of an agent successfully consuming llms.txt end-to-end beyond the probe.

   • [probe] “PROBE llms.txt: HTTP 200 at https://docs.growthbook.io/llms.txt # GrowthBook Documentation - [Docs (180 pages)](https://docs.growthbook.io/…”
   • [claimed-docs] “Connect AI tools like Cursor, VS Code, and Claude to GrowthBook using the Model Context Protocol (MCP).”
   • [claimed-docs] “The official GrowthBook MCP server is a thin bridge: it loads GrowthBook Agent Skills (workflows and guardrails) and makes authenticated cal…”
   • [probe] “official MCP server documented at https://docs.growthbook.io/integrations/mcp”

   proof ↗⚑ flag

2. ai-native userRun the product headlessly / in CI for automation

   weight 2 · round to GrowthBook
   LaunchDarkly✓fullprobed7/10

   LaunchDarkly ships an official CLI (ldcli) confirmed to run keylessly in a runtime probe, plus API access tokens, SDKs, and webhooks that together support headless/CI automation of flag management and evaluation. Missing for 10: explicit CI/CD pipeline integration examples (e.g., GitHub Actions recipes) or documented non-interactive CLI auth flows for pipelines.

   • [probe] “PROBE runtime (recorded 2026-09-05): the official LaunchDarkly CLI installed via `brew tap launchdarkly/homebrew-tap && brew install ldcli` …”
   • [claimed-docs] “LaunchDarkly CLI”
   • [claimed-docs] “API access tokens”
   • [claimed-docs] “Webhooks”
   • [claimed-docs] “30 idiomatic SDKs”

   proof ↗⚑ flag

   GrowthBook✓fullprobed8/10

   GrowthBook supports headless/CI usage via self-hosted Docker Compose deployment, a full REST API, SDK Connection Endpoints for programmatic flag evaluation, and inline server-side experiments requiring no third-party requests — all scriptable without a UI. A runtime probe confirms a real keyless self-host create-and-evaluate loop via API/docker, demonstrating genuine automation capability. missing for 10: no explicit CI/CD pipeline example or GitHub Actions integration doc, and no evidence of a dedicated CLI tool for automation workflows.

   • [claimed-docs] “Run GrowthBook on your own infrastructure with Docker Compose, including MongoDB, the API, and the front-end.”
   • [claimed-docs] “GrowthBook offers a full REST API for interacting with the application.”
   • [claimed-docs] “You can also run server-side experiments by using inline experiments directly with our SDK. This requires no 3rd party requests”
   • [claimed-docs] “The SDK Connection Endpoint provides readonly access to a subset of your feature flag data, just enough for the GrowthBook SDKs to assign va…”
   • [claimed-docs] “You can use docker compose to get started quickly... run docker compose up -d to start everything and view the app at http://localhost:3000”
   • [probe] “PROBE runtime (recorded 2026-09-05): full KEYLESS self-host roundtrip on this machine — the official growthbook/growthbook docker image boot…”

   proof ↗⚑ flag

3. ai-native userConnect an agent via an official MCP server

   weight 3 · round drawn
   LaunchDarkly✓fullprobed9/10

   LaunchDarkly documents and ships both a local npm MCP server and a hosted MCP server endpoint, and runtime probes confirm the npm server completes a full stdio initialize handshake and the hosted endpoint is live and auth-gated as documented. Missing for 10: independent third-party (non-vendor) hands-on review of the MCP server's tool functionality beyond the handshake.

   • [claimed-docs] “LaunchDarkly MCP server”
   • [claimed-docs] “Set up the LaunchDarkly hosted MCP server”
   • [probe] “official MCP server documented at https://launchdarkly.com/docs/home/getting-started/mcp”
   • [probe] “PROBE runtime (recorded 2026-09-05): the official npm MCP server (`npx -y @launchdarkly/mcp-server start --transport stdio`) completed a FUL…”
   • [probe] “PROBE runtime (recorded 2026-09-05): keyless JSON-RPC initialize POST to the documented hosted MCP endpoint https://mcp.launchdarkly.com/mcp…”

   proof ↗⚑ flag

   GrowthBook✓fullprobed9/10

   GrowthBook documents an official MCP server that bridges AI tools (Cursor, VS Code, Claude) to GrowthBook via authenticated REST calls with Agent Skills/guardrails, and a runtime probe confirms the official npm package (@growthbook/mcp) completes a full stdio initialize handshake exposing read/write meta-tools. This is strong first-party plus hands-on evidence of a working official MCP server. Missing for 10: broader independent/community corroboration beyond the single runtime probe.

   • [claimed-docs] “Connect AI tools like Cursor, VS Code, and Claude to GrowthBook using the Model Context Protocol (MCP).”
   • [claimed-docs] “The official GrowthBook MCP server is a thin bridge: it loads GrowthBook Agent Skills (workflows and guardrails) and makes authenticated cal…”
   • [probe] “official MCP server documented at https://docs.growthbook.io/integrations/mcp”
   • [probe] “PROBE runtime (recorded 2026-09-05): the official npm MCP server (`npx -y @growthbook/mcp`) completed a FULL keyless stdio initialize handsh…”

   proof ↗⚑ flag

4. ai-native userUse an official CLI

   weight 2 · round to LaunchDarkly
   LaunchDarkly✓fullprobed9/10

   LaunchDarkly ships an official CLI (`ldcli`), documented and verified via a hands-on runtime probe that installed via Homebrew and successfully ran `ldcli --version` keylessly, confirming it is real and functional. Missing for 10: independent third-party community commentary specifically evaluating the CLI's AI-native/agentic workflows.

   • [claimed-docs] “LaunchDarkly CLI”
   • [probe] “official CLI documented at https://launchdarkly.com/docs/home/getting-started/ldcli”
   • [probe] “PROBE runtime (recorded 2026-09-05): the official LaunchDarkly CLI installed via `brew tap launchdarkly/homebrew-tap && brew install ldcli` …”

   proof ↗⚑ flag

   GrowthBook—none0/10

   The evidence pack documents SDKs, REST API, self-host Docker, and an official MCP server, but there is no mention anywhere of an official standalone CLI tool for GrowthBook. Since a platform like GrowthBook could plausibly ship a CLI, the axis applies, but with no supporting evidence the verdict is none.

   ⚑ flag

5. ai-native userDrive the product through a documented public API

   weight 3 · round to GrowthBook
   LaunchDarkly✓fullprobed7/10

   LaunchDarkly documents API access tokens for programmatic access (launchdarkly-docs-37), an official CLI (launchdarkly-docs-12/probe-5) and both local and hosted MCP servers that were verified live at runtime (launchdarkly-probe-rt-1/2/3), all of which wrap the public API for automated/agentic control. Missing for 10: a discoverable OpenAPI/swagger spec (probe-3 shows 404s at standard paths) and explicit REST API reference docs in the evidence pack.

   • [claimed-docs] “API access tokens”
   • [claimed-docs] “LaunchDarkly CLI”
   • [probe] “official CLI documented at https://launchdarkly.com/docs/home/getting-started/ldcli”
   • [probe] “PROBE runtime (recorded 2026-09-05): the official LaunchDarkly CLI installed via `brew tap launchdarkly/homebrew-tap && brew install ldcli` …”
   • [probe] “PROBE runtime (recorded 2026-09-05): the official npm MCP server (`npx -y @launchdarkly/mcp-server start --transport stdio`) completed a FUL…”
   • [probe] “PROBE runtime (recorded 2026-09-05): keyless JSON-RPC initialize POST to the documented hosted MCP endpoint https://mcp.launchdarkly.com/mcp…”
   • [probe] “PROBE openapi: all candidate paths 404 (https://launchdarkly.com/openapi.json, https://launchdarkly.com/swagger.json, https://launchdarkly.c…”

   proof ↗⚑ flag

   GrowthBook✓fullprobed9/10

   GrowthBook documents a full REST API for interacting with the application (growthbook-docs-12), including a readonly SDK Connection Endpoint (growthbook-docs-20) and unauthenticated public endpoints (growthbook-docs-41), and a runtime probe confirms an actual API-driven create-and-evaluate loop (creating org, feature flag, SDK connection, and reading it back) worked end-to-end (growthbook-probe-rt-2). Missing for 10: a discoverable OpenAPI/swagger spec file (all candidate paths 404'd per growthbook-probe-2), which would make the API more machine-discoverable for agentic tooling.

   • [claimed-docs] “GrowthBook offers a full REST API for interacting with the application.”
   • [claimed-docs] “The SDK Connection Endpoint provides readonly access to a subset of your feature flag data, just enough for the GrowthBook SDKs to assign va…”
   • [claimed-docs] “They are meant to be public and do not require authentication to view.”
   • [probe] “PROBE runtime (recorded 2026-09-05): full KEYLESS self-host roundtrip on this machine — the official growthbook/growthbook docker image boot…”
   • [probe] “PROBE openapi: all candidate paths 404 (https://docs.growthbook.io/openapi.json, https://docs.growthbook.io/swagger.json, https://docs.growt…”

   proof ↗⚑ flag

6. ai-native userIssue scoped/least-privilege API credentials for an agent

   weight 2 · round to LaunchDarkly
   LaunchDarkly~partialprobed6/10

   LaunchDarkly documents API access tokens and custom role assignment, which together let a user create scoped, least-privilege credentials for any client — and the MCP server explicitly requires an API key for tool calls (agent authenticates via this token), confirming the mechanism applies to AI-agent use. However, no docs specifically address agent-oriented token types, short-lived/ephemeral credentials, or per-tool scoping tailored to autonomous agents. missing for 10: agent-specific credential type or lifecycle, fine-grained per-tool/action scoping beyond general custom roles, independent verification that scoped tokens work end-to-end with the MCP server.

   • [claimed-docs] “API access tokens”
   • [claimed-docs] “Assigning roles to members”
   • [probe] “PROBE runtime (recorded 2026-09-05): the official npm MCP server (`npx -y @launchdarkly/mcp-server start --transport stdio`) completed a FUL…”
   • [probe] “PROBE runtime (recorded 2026-09-05): keyless JSON-RPC initialize POST to the documented hosted MCP endpoint https://mcp.launchdarkly.com/mcp…”
   • [claimed-docs] “LaunchDarkly MCP server”

   proof ↗⚑ flag

   GrowthBook~partialprobed5/10

   GrowthBook documents a read-only SDK Connection Endpoint that exposes only a subset of feature-flag data (least-privilege by design) and a full REST API, plus an 'Advanced access control' / 'Advanced permissioning' feature at higher pricing tiers, and the MCP server makes authenticated REST calls on behalf of an agent. However there is no explicit documentation of an admin flow for issuing a scoped, least-privilege API credential specifically for an AI agent (e.g., role-restricted API keys, token scoping UI, or agent-specific key generation). Missing for 10: documented API-key scoping/role assignment workflow, explicit agent-credential issuance guide, and evidence of granular per-tool permission limits for MCP/agent tokens beyond generic 'advanced access control' tier feature.

   • [claimed-docs] “The SDK Connection Endpoint provides readonly access to a subset of your feature flag data, just enough for the GrowthBook SDKs to assign va…”
   • [claimed-docs] “They are meant to be public and do not require authentication to view.”
   • [claimed-docs] “GrowthBook offers a full REST API for interacting with the application.”
   • [claimed-docs] “Pro, plus: Custom environments Ramp schedules Approval workflows Advanced access control SSO & SCIM provisioning Exportable audit logs”
   • [claimed-docs] “Custom environments * Ramp schedules * Approval workflows * Advanced access control * SSO & SCIM provisioning * Exportable audit l…”
   • [claimed-docs] “The official GrowthBook MCP server is a thin bridge: it loads GrowthBook Agent Skills (workflows and guardrails) and makes authenticated cal…”
   • [probe] “PROBE runtime (recorded 2026-09-05): the official npm MCP server (`npx -y @growthbook/mcp`) completed a FULL keyless stdio initialize handsh…”

   proof ↗⚑ flag

7. ai-native userBuild against official SDKs

   weight 2 · round to GrowthBook
   LaunchDarkly✓fullprobed8/10

   LaunchDarkly documents 30 idiomatic SDKs, OpenFeature providers, and edge SDKs (e.g., Cloudflare) plus a CLI and MCP server for AI-native/agentic workflows, giving developers official building blocks across languages and AI tooling. missing for 10: independent hands-on developer reviews specifically praising SDK quality/documentation depth, and no evidence of AI-specific SDK features (e.g., LLM-context helpers) beyond general SDK/OpenFeature support.

   • [claimed-docs] “30 idiomatic SDKs”
   • [claimed-docs] “OpenFeature providers”
   • [claimed-docs] “Cloudflare SDK reference”
   • [claimed-docs] “LaunchDarkly CLI”
   • [claimed-docs] “LaunchDarkly MCP server”
   • [probe] “PROBE runtime (recorded 2026-09-05): the official LaunchDarkly CLI installed via `brew tap launchdarkly/homebrew-tap && brew install ldcli` …”
   • [probe] “PROBE runtime (recorded 2026-09-05): the official npm MCP server (`npx -y @launchdarkly/mcp-server start --transport stdio`) completed a FUL…”

   proof ↗⚑ flag

   GrowthBook✓fullcommunity9/10

   GrowthBook ships 24 official language-specific SDKs (React, Python, Android, iOS, etc.) with documented quickstart code, local evaluation, targeting, and inline experiments, and this is corroborated by community hands-on reports of fast SDK integration. missing for 10: independent benchmarking across all 24 SDKs and more recent third-party validation beyond a single HN thread.

   • [github] “24 SDKs including React, Python, Android, and iOS.”
   • [claimed-docs] “Get GrowthBook feature flags and experiments running in your app in a few lines of code, then pick a language-specific SDK.”
   • [claimed-docs] “const showNewFeature = useFeatureIsOn("new-feature");”
   • [claimed-docs] “const showNewFeature = useFeatureIsOn("new-feature"); const buttonColor = useFeatureValue("button-color", "blue");”
   • [claimed-docs] “You can also run server-side experiments by using inline experiments directly with our SDK. This requires no 3rd party requests”
   • [community] “Been using GrowthBook for a few weeks now... I was shocked at how much setup time it took to get a simple feature flag into our app with oth…”
   • [community] “At Miggos we have implemented Growth book and we got FF and A/B in our Vue frontend in a few hours. Excited to see the progress of the platf…”

   proof ↗⚑ flag

8. ai-native userSubscribe to events via webhooks

   weight 2 · round to LaunchDarkly
   LaunchDarkly✓fullclaimed8/10

   LaunchDarkly documents a dedicated Webhooks feature under infrastructure integrations, allowing subscription to flag/project events for automation, which directly supports agentic/event-driven workflows. Missing for 10: independent hands-on confirmation of webhook payload structure/reliability and no evidence of fine-grained event filtering options.

   • [claimed-docs] “Webhooks”

   proof ↗⚑ flag

   GrowthBook—none0/10

   The evidence pack covers SDKs, REST API, MCP integration, self-hosting, and experimentation features but contains no mention of webhooks or event subscription mechanisms anywhere, so there is no evidence GrowthBook supports this capability.

   ⚑ flag

Agentic features

1. ai-native userGet AI-generated insights and suggestions from my data inside the product

   weight 2 · round to GrowthBook
   LaunchDarkly—none0/10

   The evidence shows LaunchDarkly's AI-related features (AI Configs, 'AI Runs', LLM-as-judge evals) are about managing and evaluating AI/LLM application behavior via flags, not about the product itself surfacing AI-generated insights or suggestions from a user's flag/experiment/metrics data. No docs mention anomaly detection copilots, natural-language querying, or AI-generated recommendations inside the dashboard.

   • [claimed-docs] “Configure, evaluate, roll out, and observe your AI end to end”
   • [claimed-docs] “An AI Run is generated when you track an LLM request or run an online eval (LLM-as-a-judge).”
   • [claimed-docs] “Guarded rollouts”
   • [claimed-docs] “Decision making with Bayesian statistics”

   proof ↗⚑ flag

   GrowthBook~partialprobed4/10

   GrowthBook ships an official MCP server that lets AI tools like Cursor, VS Code, and Claude connect and make authenticated calls to pull experiment/feature data and get GrowthBook 'Agent Skills' workflows/guardrails, which can produce AI-generated insights and suggestions — this was verified working in a runtime probe. However, this capability lives entirely outside the product (in an external AI client) rather than as a native in-app AI insights feature; GrowthBook's own results UI only shows statistical outputs like 'Chance to Win', not LLM-generated narrative insights. Missing for 10: a native in-product AI insights/summarization feature surfaced in the GrowthBook UI itself, and evidence of automated suggestion-generation (e.g., 'this metric moved because X') without requiring an external MCP client.

   • [claimed-docs] “Connect AI tools like Cursor, VS Code, and Claude to GrowthBook using the Model Context Protocol (MCP).”
   • [claimed-docs] “The official GrowthBook MCP server is a thin bridge: it loads GrowthBook Agent Skills (workflows and guardrails) and makes authenticated cal…”
   • [probe] “PROBE runtime (recorded 2026-09-05): the official npm MCP server (`npx -y @growthbook/mcp`) completed a FULL keyless stdio initialize handsh…”
   • [claimed-docs] “Chance to Win" tells you the probability that the variation is better. Anything above 95%”

   proof ↗⚑ flag

2. ai-native userSet up automations that run autonomously in the background

   weight 2 · round to LaunchDarkly
   LaunchDarkly~partialprobed5/10

   LaunchDarkly supports autonomous background automations like scheduled flag changes and guarded rollouts that automatically monitor metrics and roll back without human intervention, plus webhooks that can trigger downstream automation, and its MCP server lets an AI agent configure these. But there's no evidence of a general-purpose agentic automation/orchestration capability beyond release-management rules. Missing for 10: evidence of AI agents autonomously creating/managing ongoing automations via MCP (only initial handshake is verified), broader agentic workflow orchestration beyond flag/rollout rules, and independent confirmation these autonomous rollbacks work reliably in practice.

   • [claimed-docs] “Scheduled flag changes”
   • [claimed-docs] “Guarded rollouts”
   • [claimed-docs] “Webhooks”
   • [claimed-docs] “LaunchDarkly MCP server”
   • [probe] “PROBE runtime (recorded 2026-09-05): the official npm MCP server (`npx -y @launchdarkly/mcp-server start --transport stdio`) completed a FUL…”

   proof ↗⚑ flag

   GrowthBook~partialprobed4/10

   GrowthBook has some automated background behaviors (automatic stale flag detection, ramp schedules, and 'safe rollouts' that presumably auto-guard rollouts), and its MCP server lets AI tools make authenticated calls to the product, but these are on-demand agent actions or passive detection features, not evidence of an AI agent configuring a persistent job that runs autonomously in the background on a schedule/trigger without human involvement. Missing for 10: a documented scheduler/trigger system for AI-configured autonomous jobs, evidence of MCP-driven automations persisting and executing unattended, and independent confirmation of any autonomous background workflow.

   • [claimed-docs] “GrowthBook can automatically detect these flags so you can clean them up.”
   • [claimed-docs] “How rules override a feature's default value — targeting users, rolling out gradually, experimenting — and how to scope, ramp, and schedule …”
   • [claimed-docs] “Pro, plus: Custom environments Ramp schedules Approval workflows Advanced access control SSO & SCIM provisioning Exportable audit logs”
   • [claimed-docs] “Custom environments * Ramp schedules * Approval workflows * Advanced access control * SSO & SCIM provisioning * Exportable audit l…”
   • [claimed-docs] “Starter, plus: Visual editor Multi-arm bandits Safe rollouts Customizable dashboards Advanced permissioning Power calculator Premium support”
   • [claimed-docs] “The official GrowthBook MCP server is a thin bridge: it loads GrowthBook Agent Skills (workflows and guardrails) and makes authenticated cal…”
   • [probe] “PROBE runtime (recorded 2026-09-05): the official npm MCP server (`npx -y @growthbook/mcp`) completed a FULL keyless stdio initialize handsh…”

   proof ↗⚑ flag

3. ai-native userDelegate tasks to a built-in AI assistant inside the product

   weight 3 · round drawn
   LaunchDarkly—none0/10

   LaunchDarkly's evidence shows AI-related feature-flag tooling (AI Runs, LLM evaluation flags) and an MCP server that lets external AI agents call into LaunchDarkly as a tool provider, but there is no evidence of a built-in AI assistant inside the product itself that a user can delegate tasks to.

   • [claimed-docs] “Configure, evaluate, roll out, and observe your AI end to end”
   • [claimed-docs] “An AI Run is generated when you track an LLM request or run an online eval (LLM-as-a-judge).”
   • [claimed-docs] “LaunchDarkly MCP server”
   • [claimed-docs] “Set up the LaunchDarkly hosted MCP server”

   proof ↗⚑ flag

   GrowthBook—none0/10

   GrowthBook documents an MCP server that lets external AI tools (Cursor, VS Code, Claude) connect to and query GrowthBook — this is the opposite direction (GrowthBook as a data/tool source for external assistants), not a built-in AI assistant living inside the GrowthBook product itself. No evidence anywhere in the pack shows an in-app chat/assistant feature for delegating tasks within GrowthBook's UI.

   • [claimed-docs] “Connect AI tools like Cursor, VS Code, and Claude to GrowthBook using the Model Context Protocol (MCP).”
   • [claimed-docs] “The official GrowthBook MCP server is a thin bridge: it loads GrowthBook Agent Skills (workflows and guardrails) and makes authenticated cal…”
   • [probe] “official MCP server documented at https://docs.growthbook.io/integrations/mcp”
   • [probe] “PROBE runtime (recorded 2026-09-05): the official npm MCP server (`npx -y @growthbook/mcp`) completed a FULL keyless stdio initialize handsh…”

   proof ↗⚑ flag

4. ai-native userOperate the product with natural-language commands

   weight 2 · round to GrowthBook
   LaunchDarkly✓fullprobed7/10

   LaunchDarkly ships an official MCP server (local npm package and hosted endpoint) that exposes flag/targeting/rollout operations to AI agents, letting an AI-native user issue natural-language commands through an MCP-compatible client; runtime probes confirm the local server completes a full stdio handshake and the hosted endpoint is live and auth-gated as documented. Missing for 10: no evidence of actual authenticated tool-call transcripts showing natural-language commands successfully executing flag changes, and no independent/community corroboration of the MCP workflow in practice.

   • [claimed-docs] “LaunchDarkly MCP server”
   • [claimed-docs] “Set up the LaunchDarkly hosted MCP server”
   • [probe] “official MCP server documented at https://launchdarkly.com/docs/home/getting-started/mcp”
   • [probe] “PROBE runtime (recorded 2026-09-05): the official npm MCP server (`npx -y @launchdarkly/mcp-server start --transport stdio`) completed a FUL…”
   • [probe] “PROBE runtime (recorded 2026-09-05): keyless JSON-RPC initialize POST to the documented hosted MCP endpoint https://mcp.launchdarkly.com/mcp…”

   proof ↗⚑ flag

   GrowthBook✓fullprobed8/10

   GrowthBook ships an official MCP server that bridges AI tools like Cursor, VS Code, and Claude to GrowthBook, loading 'Agent Skills' (workflows/guardrails) and making authenticated REST calls so users can operate flags/experiments via natural language in an AI client. A runtime probe confirms the official npm MCP server completes a full keyless handshake exposing read/write meta-tools covering the whole API, corroborating the docs' claims beyond marketing copy. Missing for 10: an end-to-end example of an actual natural-language command executing a real write action (only the handshake/tool-discovery was verified, not a full command-to-result flow) and independent (non-vendor) hands-on confirmation.

   • [claimed-docs] “Connect AI tools like Cursor, VS Code, and Claude to GrowthBook using the Model Context Protocol (MCP).”
   • [claimed-docs] “The official GrowthBook MCP server is a thin bridge: it loads GrowthBook Agent Skills (workflows and guardrails) and makes authenticated cal…”
   • [probe] “official MCP server documented at https://docs.growthbook.io/integrations/mcp”
   • [probe] “PROBE runtime (recorded 2026-09-05): the official npm MCP server (`npx -y @growthbook/mcp`) completed a FULL keyless stdio initialize handsh…”

   proof ↗⚑ flag

Api quality

1. ai-native userExplore an interactive API reference with runnable examples

   weight 2 · round drawn
   LaunchDarkly—none0/10

   The evidence pack shows only static docs pages and confirms no OpenAPI/swagger spec discoverable (probe-3 all 404s) and no llms.txt/docs.md; nothing indicates an interactive, runnable API reference/playground exists.

   • [probe] “PROBE openapi: all candidate paths 404 (https://launchdarkly.com/openapi.json, https://launchdarkly.com/swagger.json, https://launchdarkly.c…”
   • [probe] “PROBE llms.txt: HTTP 404 at https://launchdarkly.com/llms.txt”
   • [probe] “PROBE docs-md: HTTP 404 at https://launchdarkly.com/docs.md”

   proof ↗⚑ flag

   GrowthBook—none0/10

   GrowthBook documents a REST API (growthbook-docs-12) but there is no evidence of an interactive, runnable API reference (e.g., Swagger/OpenAPI try-it console) — a direct probe for openapi/swagger specs returned 404 on all candidate paths, and no docs mention runnable code samples in an API explorer.

   • [claimed-docs] “GrowthBook offers a full REST API for interacting with the application.”
   • [probe] “PROBE openapi: all candidate paths 404 (https://docs.growthbook.io/openapi.json, https://docs.growthbook.io/swagger.json, https://docs.growt…”

   proof ↗⚑ flag

2. ai-native userDownload a machine-readable API spec (OpenAPI or equivalent)

   weight 2 · round drawn
   LaunchDarkly—none0/10

   The evidence pack shows explicit probes for an OpenAPI/swagger spec at LaunchDarkly's standard candidate paths all returning 404, and no other citation surfaces a downloadable machine-readable API spec (only human-readable API access token docs are referenced). Despite LaunchDarkly having a public REST API in reality, this evidence pack contains no proof of an accessible OpenAPI spec.

   • [probe] “PROBE openapi: all candidate paths 404 (https://launchdarkly.com/openapi.json, https://launchdarkly.com/swagger.json, https://launchdarkly.c…”
   • [claimed-docs] “API access tokens”

   proof ↗⚑ flag

   GrowthBook—none0/10

   While GrowthBook documents a full REST API (growthbook-docs-12) and an SDK Connection Endpoint (growthbook-docs-20), the evidence pack shows an explicit probe for OpenAPI/Swagger spec files at expected paths that all returned 404, and no other evidence surfaces a downloadable OpenAPI or machine-readable spec.

   • [claimed-docs] “GrowthBook offers a full REST API for interacting with the application.”
   • [probe] “PROBE openapi: all candidate paths 404 (https://docs.growthbook.io/openapi.json, https://docs.growthbook.io/swagger.json, https://docs.growt…”

   proof ↗⚑ flag

3. ai-native userTest against a sandbox environment without touching production data

   weight 1 · round to GrowthBook
   LaunchDarkly✓fullclaimed7/10

   LaunchDarkly's Environments feature explicitly separates test/staging from production, letting users evaluate and target flags in a non-production environment, and 'Testing changes to flag targeting' documents workflows to validate rule changes safely before they affect live traffic. Missing for 10: no explicit 'sandbox for AI-native/agent testing' framing, no independent/hands-on confirmation that environment data is fully isolated from production telemetry.

   • [claimed-docs] “Environments”
   • [claimed-docs] “Testing changes to flag targeting”
   • [claimed-docs] “Approvals”
   • [claimed-docs] “Change history”

   proof ↗⚑ flag

   GrowthBook✓fullprobed8/10

   GrowthBook supports separate environments (dev/staging/production) for feature flags, self-hosting via Docker Compose for a fully isolated instance, and a verified runtime probe confirms a keyless self-host roundtrip creating and evaluating flags without any production account/license — enabling true sandbox testing. missing for 10: no explicit vendor guidance on a dedicated 'sandbox mode' distinct from staging environments, and no independent (non-vendor) confirmation of environment data isolation guarantees.

   • [claimed-docs] “Use environments such as dev, staging, and production to enable or disable feature flags independently.”
   • [claimed-docs] “Run GrowthBook on your own infrastructure with Docker Compose, including MongoDB, the API, and the front-end.”
   • [claimed-docs] “You can use docker compose to get started quickly... run docker compose up -d to start everything and view the app at http://localhost:3000”
   • [probe] “PROBE runtime (recorded 2026-09-05): full KEYLESS self-host roundtrip on this machine — the official growthbook/growthbook docker image boot…”

   proof ↗⚑ flag

4. ai-native userRely on versioned APIs with a documented deprecation policy

   weight 2 · round drawn
   LaunchDarkly—none0/10

   Evidence shows API access tokens and CLI/MCP tooling but no documentation of API versioning scheme or a deprecation policy; OpenAPI probes returned 404s, providing no confirmation of a discoverable, versioned API spec.

   • [claimed-docs] “API access tokens”
   • [probe] “PROBE openapi: all candidate paths 404 (https://launchdarkly.com/openapi.json, https://launchdarkly.com/swagger.json, https://launchdarkly.c…”

   proof ↗⚑ flag

   GrowthBook—none0/10

   GrowthBook documents a REST API and SDK Connection Endpoint, but no evidence pack item describes API versioning scheme or a documented deprecation policy; the openapi/swagger probe returned 404s, suggesting no formal API spec is even discoverable. Missing for 10: any mention of API version numbers, backward-compatibility guarantees, or a deprecation/sunset policy.

   • [claimed-docs] “GrowthBook offers a full REST API for interacting with the application.”
   • [claimed-docs] “The SDK Connection Endpoint provides readonly access to a subset of your feature flag data, just enough for the GrowthBook SDKs to assign va…”
   • [probe] “PROBE openapi: all candidate paths 404 (https://docs.growthbook.io/openapi.json, https://docs.growthbook.io/swagger.json, https://docs.growt…”

   proof ↗⚑ flag

1. ai-native userPerform bulk operations across many items at once

   weight 2 · round drawn
   LaunchDarkly~partialprobed5/10

   LaunchDarkly exposes a CLI, REST API tokens, and an MCP server that could be scripted by an AI-native user to act on many flags/segments programmatically, but no documented UI or API feature explicitly named 'bulk operations' (e.g., bulk edit/bulk archive across flags) appears in the evidence pack. Missing for 10: explicit bulk-edit/bulk-update API endpoints or UI, documentation of batch operations across contexts/segments, and independent confirmation of bulk-scale usage.

   • [claimed-docs] “LaunchDarkly CLI”
   • [claimed-docs] “API access tokens”
   • [probe] “official MCP server documented at https://launchdarkly.com/docs/home/getting-started/mcp”
   • [probe] “official CLI documented at https://launchdarkly.com/docs/home/getting-started/ldcli”
   • [probe] “PROBE runtime (recorded 2026-09-05): the official LaunchDarkly CLI installed via `brew tap launchdarkly/homebrew-tap && brew install ldcli` …”
   • [probe] “PROBE runtime (recorded 2026-09-05): the official npm MCP server (`npx -y @launchdarkly/mcp-server start --transport stdio`) completed a FUL…”

   proof ↗⚑ flag

   GrowthBook~partialprobed5/10

   GrowthBook exposes a full REST API (growthbook-docs-12) and an official MCP server that bridges AI tools to authenticated REST read/write calls across the whole API (growthbook-docs-21, growthbook-probe-rt-1), which could let an AI agent script operations across many flags/experiments programmatically. However, there is no documented bulk-specific endpoint, batch-update tool, or example of performing many-item operations in one call. Missing for 10: explicit bulk-create/bulk-update API endpoints, documented batch operation examples, and evidence of the MCP server or API actually handling many items in a single request rather than one-at-a-time calls.

   • [claimed-docs] “GrowthBook offers a full REST API for interacting with the application.”
   • [claimed-docs] “The official GrowthBook MCP server is a thin bridge: it loads GrowthBook Agent Skills (workflows and guardrails) and makes authenticated cal…”
   • [probe] “PROBE runtime (recorded 2026-09-05): the official npm MCP server (`npx -y @growthbook/mcp`) completed a FULL keyless stdio initialize handsh…”

   proof ↗⚑ flag

2. ai-native userDefine rules that trigger actions automatically on events

   weight 3 · round to LaunchDarkly
   LaunchDarkly✓fullclaimed7/10

   LaunchDarkly ships rule-based targeting (custom rules, adaptive triggers), scheduled flag changes, and webhooks that let flag state changes or conditions automatically trigger downstream actions/integrations — directly matching the 'rules trigger actions on events' story. Missing for 10: independent/hands-on verification of adaptive-trigger and webhook automation behavior in practice, and clearer documentation of arbitrary event-driven (vs. flag-state) triggers.

   • [claimed-docs] “Adaptive triggers”
   • [claimed-docs] “Custom rules”
   • [claimed-docs] “Scheduled flag changes”
   • [claimed-docs] “Webhooks”
   • [claimed-docs] “Targeting rules”

   proof ↗⚑ flag

   GrowthBook~partialclaimed5/10

   GrowthBook lets users define feature-flag 'rules' that automatically apply based on targeting conditions, scheduling, and ramping (docs-22, docs-28, docs-2), and has automated stale-flag detection (docs-4) — these are rule-based automations tied to feature evaluation. However, this falls short of a general 'on event X, trigger action Y' automation engine; there's no evidence of webhook-driven arbitrary actions, external event ingestion, or an automation/workflow layer beyond flag targeting and rollout scheduling. Missing for 10: generic event-trigger/webhook automation outside feature-flag context, evidence of arbitrary action execution on external events, and independent corroboration of rule automation depth.

   • [claimed-docs] “How rules override a feature's default value — targeting users, rolling out gradually, experimenting — and how to scope, ramp, and schedule …”
   • [claimed-docs] “Add targeting conditions to a rule to control which values a user sees.”
   • [claimed-docs] “GrowthBook can automatically detect these flags so you can clean them up.”
   • [claimed-docs] “Add targeting conditions to a rule to control which values a user sees. Conditions are evaluated against attributes you pass into the SDK”

   proof ↗⚑ flag

3. ai-native userSchedule recurring jobs or workflows

   weight 2 · round drawn
   LaunchDarkly—none0/10

   LaunchDarkly documents one-time 'Scheduled flag changes' (docs-35) but there is no evidence of recurring/cron-style job or workflow scheduling; its CLI, API, and MCP server (docs-10/11/12) could be used by external automation but no first-party recurring scheduler is documented.

   • [claimed-docs] “Scheduled flag changes”
   • [claimed-docs] “LaunchDarkly CLI”
   • [claimed-docs] “LaunchDarkly MCP server”

   proof ↗⚑ flag

   GrowthBook—none0/10

   Evidence only shows one-time scheduling of feature-flag rule start/stop dates and 'ramp schedules' for gradual rollouts (growthbook-docs-22, growthbook-docs-37) — nothing resembling recurring cron-like jobs or workflow automation that an AI-native user could schedule to run repeatedly. No mention of scheduled reports, recurring experiment analyses, or workflow automation triggers.

   • [claimed-docs] “How rules override a feature's default value — targeting users, rolling out gradually, experimenting — and how to scope, ramp, and schedule …”
   • [claimed-docs] “Custom environments * Ramp schedules * Approval workflows * Advanced access control * SSO & SCIM provisioning * Exportable audit l…”
   • [claimed-docs] “Pro, plus: Custom environments Ramp schedules Approval workflows Advanced access control SSO & SCIM provisioning Exportable audit logs”

   proof ↗⚑ flag

4. ai-native userVersion, review, and roll back my automations

   weight 1 · round to GrowthBook
   LaunchDarkly~partialclaimed5/10

   LaunchDarkly's flags (its closest analog to 'automations') have change history, approvals, scheduled changes, and rollback via targeting/rollout controls, which cover versioning and review workflows for flag changes. However, this is framed around feature flags/rollouts, not AI agent automations specifically, and there's no evidence of a dedicated automation-versioning UI, diff/review workflow for AI-agent-driven changes, or audit trail tied to AI-native usage. missing for 10: explicit versioning/rollback UI for AI-generated automations, review/approval workflow tied to AI agent changes, independent confirmation that change history supports full rollback of complex rule sets.

   • [claimed-docs] “Approvals”
   • [claimed-docs] “Change history”
   • [claimed-docs] “Scheduled flag changes”
   • [claimed-docs] “Creating and managing progressive rollouts”
   • [claimed-docs] “Guarded rollouts”

   proof ↗⚑ flag

   GrowthBook~partialcommunity6/10

   GrowthBook documents draft feature revisions, request-review, merge-conflict resolution, and required approval before publishing (docs-5, docs-22), and lists approval workflows plus exportable audit logs as paid-tier features (docs-37), which together enable versioning and review of flag/automation changes. However, no explicit rollback mechanism is documented and there's no independent/hands-on corroboration of the revision-history or rollback UI; a community comment even suggests config-as-code/git-based versioning isn't natively supported (growthbook-comm-11). Missing for 10: explicit rollback documentation, hands-on verification of revision history UI, and clarity on git-based config versioning.

   • [claimed-docs] “Draft feature revisions, request review, resolve merge conflicts, and require approval before publishing.”
   • [claimed-docs] “How rules override a feature's default value — targeting users, rolling out gradually, experimenting — and how to scope, ramp, and schedule …”
   • [claimed-docs] “Custom environments * Ramp schedules * Approval workflows * Advanced access control * SSO & SCIM provisioning * Exportable audit l…”
   • [community] “From the looks of it, it doesn't look like the configuration can be stored in the code repository itself. This is one of the key things to d…”

   proof ↗⚑ flag

Proxy

1. platform engineerRun a relay/edge proxy so flags stay served when the vendor is unreachable and SDK traffic stays inside my network

   weight 1 · round to GrowthBook
   LaunchDarkly✓fullcommunity8/10

   LaunchDarkly documents the Relay Proxy as a self-hosted Go application that runs on customer infrastructure to serve flags and cache SDK connections, explicitly for resiliency, and this is independently corroborated by community commentary describing it as open source and used 'for extra resiliency' when LaunchDarkly's servers are unreachable, keeping SDK traffic within the customer's network. Missing for 10: independent hands-on verification of offline-serving behavior during vendor outage, and explicit first-party detail on network-isolation guarantees beyond general resiliency claims.

   • [claimed-docs] “The LaunchDarkly Relay Proxy is a small Go application that runs on your own infrastructure.”
   • [claimed-docs] “The Relay Proxy”
   • [community] “LaunchDarkly clients download all flags in the project on connection, then stream changes. ... LaunchDarkly has an open source 'Relay' you c…”

   proof ↗⚑ flag

   GrowthBook✓fullcommunity9/10

   GrowthBook explicitly documents a self-hostable Proxy that runs next to your app to cache SDK payloads, reduce latency, and keep serving flags if the API is down, directly matching the relay/edge proxy story, and this sits alongside full self-host Docker Compose deployment keeping SDK traffic in-network. Community evidence corroborates that evaluation happens locally within SDKs via a cacheable JSON payload, reinforcing offline resilience. Missing for 10: independent hands-on report specifically of the Proxy component (vs. general self-host) failing over during an outage.

   • [claimed-docs] “Run the GrowthBook Proxy next to your app to cache SDK payloads, reduce latency, and keep serving flags if the API is down.”
   • [claimed-docs] “Run GrowthBook on your own infrastructure with Docker Compose, including MongoDB, the API, and the front-end.”
   • [claimed-docs] “You can use docker compose to get started quickly... run docker compose up -d to start everything and view the app at http://localhost:3000”
   • [community] “With GrowthBook, there's a single cacheable JSON definition of all of the features and then evaluation happens locally within the SDKs... Th…”
   • [claimed-docs] “The SDK Connection Endpoint provides readonly access to a subset of your feature flag data, just enough for the GrowthBook SDKs to assign va…”

   proof ↗⚑ flag

Self host

1. platform engineerSelf-host the full flag platform from an open-source distribution, keeping evaluation data on my infrastructure

   weight 3 · round to GrowthBook
   LaunchDarkly—none0/10

   LaunchDarkly's evidence only describes an open-source Relay Proxy that caches/streams flag data from the LaunchDarkly cloud for resiliency and low latency, not a full self-hostable flag-management platform; the control plane, rules engine, and dashboard remain SaaS-hosted, and community comments explicitly flag this as a third-party critical-path dependency rather than a self-hosted deployment (launchdarkly-docs-13, launchdarkly-comm-2, launchdarkly-comm-6). No evidence shows an OSS distribution of the full platform keeping evaluation data entirely on customer infrastructure.

   • [claimed-docs] “The LaunchDarkly Relay Proxy is a small Go application that runs on your own infrastructure.”
   • [claimed-docs] “The Relay Proxy”
   • [community] “LaunchDarkly clients download all flags in the project on connection, then stream changes. ... LaunchDarkly has an open source 'Relay' you c…”
   • [community] “It's a pretty question[able] architecture that outsources feature flags to a third party. You're creating a critical path hard dependency an…”

   proof ↗⚑ flag

   GrowthBook✓fullprobed9/10

   GrowthBook documents a self-host distribution via Docker Compose (MongoDB, API, front-end) and this is corroborated by a hands-on probe showing a full keyless self-host boot, feature creation, and SDK payload evaluation working end-to-end on local infrastructure, plus community confirmation of production self-hosted usage. missing for 10: independent third-party audit of data residency guarantees beyond vendor docs/probe.

   • [claimed-docs] “Run GrowthBook on your own infrastructure with Docker Compose, including MongoDB, the API, and the front-end.”
   • [claimed-docs] “You can use docker compose to get started quickly... run docker compose up -d to start everything and view the app at http://localhost:3000”
   • [probe] “PROBE runtime (recorded 2026-09-05): full KEYLESS self-host roundtrip on this machine — the official growthbook/growthbook docker image boot…”
   • [community] “Growthbook is awesome! We have used the self-hosted version a few months in my company now, and are very happy about it. Talented people beh…”

   proof ↗⚑ flag

Agent ops

1. ai agentRead experiment configurations and results programmatically to summarize outcomes and recommend ship/rollback decisions

   weight 1 · round to GrowthBook
   LaunchDarkly~partialprobed5/10

   LaunchDarkly documents an Experimentation feature with Bayesian statistical analysis, exposes API access tokens, an official CLI, and a hosted/local MCP server that an AI agent could use to read flag/experiment configuration data programmatically. However, no evidence explicitly confirms the MCP server or API exposes experiment results/metrics data (vs. just flag targeting) or that an agent has been shown reading such data to produce ship/rollback recommendations. Missing for 10: explicit documentation of an experiment-results API/MCP tool, and a hands-on example of an agent summarizing experiment outcomes or recommending ship/rollback decisions.

   • [claimed-docs] “Experimentation”
   • [claimed-docs] “Decision making with Bayesian statistics”
   • [claimed-docs] “API access tokens”
   • [claimed-docs] “LaunchDarkly MCP server”
   • [probe] “official MCP server documented at https://launchdarkly.com/docs/home/getting-started/mcp”
   • [probe] “PROBE runtime (recorded 2026-09-05): the official npm MCP server (`npx -y @launchdarkly/mcp-server start --transport stdio`) completed a FUL…”

   proof ↗⚑ flag

   GrowthBook✓fullprobed8/10

   GrowthBook exposes a full REST API and an official MCP server that makes authenticated calls to that API, explicitly loading 'Agent Skills (workflows and guardrails)' so AI tools can query feature/experiment data — confirmed working end-to-end in a runtime probe. Experiment results expose statistical decision signals like 'Chance to Win' (>95% highlighted as a clear winner) that directly support ship/rollback recommendations, and metrics/fact-table docs show how results are computed from stored data for programmatic reading. Missing for 10: a concrete worked example of an agent actually generating a ship/rollback summary via MCP, and independent (non-vendor) confirmation of recommendation quality.

   • [claimed-docs] “GrowthBook offers a full REST API for interacting with the application.”
   • [claimed-docs] “Connect AI tools like Cursor, VS Code, and Claude to GrowthBook using the Model Context Protocol (MCP).”
   • [claimed-docs] “The official GrowthBook MCP server is a thin bridge: it loads GrowthBook Agent Skills (workflows and guardrails) and makes authenticated cal…”
   • [claimed-docs] “Chance to Win" tells you the probability that the variation is better. Anything above 95%”
   • [claimed-docs] “Chance to Win tells you the probability that the variation is better. Anything above 95% ... is highlighted green indicating a very clear wi…”
   • [probe] “PROBE runtime (recorded 2026-09-05): the official npm MCP server (`npx -y @growthbook/mcp`) completed a FULL keyless stdio initialize handsh…”
   • [probe] “official MCP server documented at https://docs.growthbook.io/integrations/mcp”

   proof ↗⚑ flag

Analysis

1. product managerTrust a documented statistics engine (Bayesian or frequentist, with variance-reduction options) behind experiment results

   weight 2 · round to GrowthBook
   LaunchDarkly~partialclaimed6/10

   LaunchDarkly documents a Bayesian statistics engine for experiment results and a dedicated experimentation/metrics framework, giving PMs a documented methodology behind results. However, the evidence pack shows no explicit mention of a frequentist alternative or variance-reduction techniques (e.g., CUPED), and there's no independent/hands-on corroboration of the engine's real-world reliability. missing for 10: frequentist option documentation, variance-reduction feature details, independent validation of statistical results.

   • [claimed-docs] “Decision making with Bayesian statistics”
   • [claimed-docs] “Experimentation”
   • [claimed-docs] “Choosing a metric type”

   proof ↗⚑ flag

   GrowthBook✓fullcommunity9/10

   GrowthBook documents both Bayesian and Frequentist statistics engines with explicit UI outputs like 'Chance to Win', and its GitHub/community materials detail variance-reduction and rigor features (CUPED, Sequential testing, Post-Stratification, SRM checks) plus rationale for choosing Bayesian to mitigate peeking/Type I error. This is corroborated by both official docs and first-party community explanation, giving a PM strong documented trust in the underlying stats methodology; missing for 10: independent third-party statistical audit or academic validation of the engine's correctness.

   • [claimed-docs] “What you will see will depend a little bit on whether you are using our Bayesian or our Frequentist engine”
   • [claimed-docs] “Chance to Win" tells you the probability that the variation is better. Anything above 95%”
   • [claimed-docs] “Chance to Win tells you the probability that the variation is better. Anything above 95% ... is highlighted green indicating a very clear wi…”
   • [github] “World class experiment stats engine (CUPED, Sequential, Bayesian, Post-Strat, Bandits, SRM checks).”
   • [github] “🆎 World class experiment stats engine (CUPED, Sequential, Bayesian, Post-Strat, Bandits, SRM checks).”
   • [community] “We use a Bayesian statistics engine which is much more immune to peeking problems and Type I errors than frequentist approaches.”

   proof ↗⚑ flag

Experiments

1. product managerRun A/B and multivariate experiments on flags and see which variation wins on my metrics

   weight 3 · round drawn
   LaunchDarkly✓fullcommunity8/10

   LaunchDarkly's docs explicitly cover A/B Tests and Experiments, metric type selection, and Bayesian statistics for deciding a winning variation, directly matching the PM story of running experiments on flags and seeing which variation wins on metrics. Community evidence corroborates multivariate flags and rollout targeting as real, used features. Missing for 10: independent hands-on account of running a full experiment and reading results dashboards, and more detail on multivariate (not just binary) experiment support.

   • [claimed-docs] “A/B Tests and Experiments”
   • [claimed-docs] “Experimentation”
   • [claimed-docs] “Choosing a metric type”
   • [claimed-docs] “Decision making with Bayesian statistics”
   • [community] “Key benefits of LaunchDarkly: rules-based targeting so flags can be set at group level, custom rollout by % of users or attributes, centrali…”

   proof ↗⚑ flag

   GrowthBook✓fullcommunity8/10

   GrowthBook natively supports A/B and multivariate experiments attached to feature flags, with statistical engines (Bayesian/Frequentist, CUPED, Sequential, Bandits), templates for goal/secondary/guardrail metrics, and a results view showing 'Chance to Win' per variation to determine winners. Community reports corroborate quick setup and real-world use of feature flags plus A/B testing, though some note a gap between flags and full analytics requiring warehouse/metrics setup. missing for 10: independent hands-on validation of multivariate (>2 arm) experiment results and winner-declaration workflow specifically, beyond docs and community anecdotes.

   • [claimed-docs] “GrowthBook has a few different ways to run experiments or AB tests depending on your needs.”
   • [claimed-docs] “What you will see will depend a little bit on whether you are using our Bayesian or our Frequentist engine”
   • [claimed-docs] “Selecting a Template in the creation modal pre-fills settings like the data source, goal/secondary/guardrail metrics, targeting, and traffic”
   • [claimed-docs] “Define Fact Tables and metrics so GrowthBook can compute experiment results from your warehouse.”
   • [claimed-docs] “Chance to Win" tells you the probability that the variation is better. Anything above 95%”
   • [claimed-docs] “Run A/B tests with feature flags, the Visual Editor, URL redirects, or your own assignment, then analyze results in GrowthBook.”
   • [claimed-docs] “you can add an experiment rule to a feature that will randomly assign the users based on some hashing attribute into one of your experiment …”
   • [github] “World class experiment stats engine (CUPED, Sequential, Bayesian, Post-Strat, Bandits, SRM checks).”
   • [community] “With GrowthBook, there's a single cacheable JSON definition of all of the features and then evaluation happens locally within the SDKs... Th…”
   • [community] “We use a Bayesian statistics engine which is much more immune to peeking problems and Type I errors than frequentist approaches.”

   proof ↗⚑ flag

Metrics

1. product managerDefine experiment metrics from my own data — warehouse tables or ingested events — instead of a black-box metric store

   weight 2 · round to GrowthBook
   LaunchDarkly—none0/10

   LaunchDarkly's experimentation docs (metrics, Bayesian stats, A/B tests) describe metrics defined via LaunchDarkly's own event tracking and metric store; there is no evidence of connecting to a customer's data warehouse or ingesting external event tables to define metrics.

   • [claimed-docs] “Choosing a metric type”
   • [claimed-docs] “Experimentation”
   • [claimed-docs] “Decision making with Bayesian statistics”
   • [claimed-docs] “A/B Tests and Experiments”

   proof ↗⚑ flag

   GrowthBook✓fullcommunity8/10

   GrowthBook explicitly lets teams define Fact Tables and metrics computed directly from their own warehouse (BigQuery, Snowflake, Databricks, etc.) rather than relying on a proprietary metric store, and independent community commentary confirms this warehouse-native design as a key differentiator from black-box tools like Optimizely. missing for 10: hands-on PM-level walkthrough of defining a metric from ingested event data (vs. warehouse tables) and any independent report specifically validating metric-definition UX.

   • [claimed-docs] “Define Fact Tables and metrics so GrowthBook can compute experiment results from your warehouse.”
   • [github] “Warehouse Native. Query 11 data sources including BigQuery, Snowflake, and Databricks.”
   • [github] “❄️ Warehouse Native. Query 11 data sources including BigQuery, Snowflake, and Databricks.”
   • [community] “Re: Eppo - they're similar in setup but Growthbook is open source with an optional cloud SaaS offering. Optimizely is a walled garden, you n…”
   • [community] “a guide in your docs on how to use this with Matomo would go far. Growthbook provides Feature Flag and A/B Testing, but you have to bring yo…”

   proof ↗⚑ flag

Environments

1. developerManage separate environments (dev/staging/prod) with independent flag states and scoped SDK keys

   weight 2 · round to LaunchDarkly
   LaunchDarkly✓fullcommunity9/10

   LaunchDarkly's docs explicitly cover Environments as a first-class construct with independent flag states per environment, and API access tokens/roles docs indicate scoped SDK keys per environment; community evidence corroborates the SDK/streaming architecture supporting per-environment keys. Missing for 10: no independent hands-on confirmation of scoped SDK key isolation across environments beyond docs.

   • [claimed-docs] “Environments”
   • [claimed-docs] “API access tokens”
   • [claimed-docs] “Assigning roles to members”
   • [community] “Key benefits of LaunchDarkly: rules-based targeting so flags can be set at group level, custom rollout by % of users or attributes, centrali…”

   proof ↗⚑ flag

   GrowthBook✓fullprobed8/10

   Docs explicitly describe dedicated environments (dev/staging/production) with independently toggled flag states [growthbook-docs-3], and the SDK Connection Endpoint provides scoped, readonly keys per environment for SDKs [growthbook-docs-20], with runtime probes confirming a real create-flag/scoped-key/evaluate loop [growthbook-probe-rt-2]. Custom environments are also called out as a paid-tier feature, confirming the capability exists in-product [growthbook-docs-16, growthbook-docs-37]. Missing for 10: independent/community verification specifically of environment-scoped SDK key isolation (evidence is docs/probe only, not third-party confirmation) and more detail on default-environment limits on lower tiers.

   • [claimed-docs] “Use environments such as dev, staging, and production to enable or disable feature flags independently.”
   • [claimed-docs] “The SDK Connection Endpoint provides readonly access to a subset of your feature flag data, just enough for the GrowthBook SDKs to assign va…”
   • [claimed-docs] “Pro, plus: Custom environments Ramp schedules Approval workflows Advanced access control SSO & SCIM provisioning Exportable audit logs”
   • [claimed-docs] “Custom environments * Ramp schedules * Approval workflows * Advanced access control * SSO & SCIM provisioning * Exportable audit l…”
   • [probe] “PROBE runtime (recorded 2026-09-05): full KEYLESS self-host roundtrip on this machine — the official growthbook/growthbook docker image boot…”

   proof ↗⚑ flag

Flags

1. developerCreate a feature flag and toggle it live in production within minutes of signing up

   weight 3 · round to GrowthBook
   LaunchDarkly✓fullcommunity7/10

   LaunchDarkly's docs explicitly cover creating flags, targeting rules, percentage rollouts, and environments—the core primitives needed to create and toggle a flag live in production—and community accounts confirm real-world use of rules-based targeting and rollout by percentage/attributes with minimal latency. However, missing for 10: a documented quickstart/time-to-first-flag walkthrough proving the 'minutes after signup' claim, and independent hands-on confirmation of onboarding speed (one community comment even calls the product generally 'a pain to use').

   • [claimed-docs] “Create flags”
   • [claimed-docs] “Percentage rollouts”
   • [claimed-docs] “Environments”
   • [claimed-docs] “Unlimited feature flags”
   • [community] “Key benefits of LaunchDarkly: rules-based targeting so flags can be set at group level, custom rollout by % of users or attributes, centrali…”
   • [community] “We've used LaunchDarkly (LD) for about a year now. We previously used our own Feature Flag library built by our CTO. LD provided a lot more,…”
   • [community] “I have used LaunchDarkly in the past and yeah it's kind of a pain to use.”

   proof ↗⚑ flag

   GrowthBook✓fullprobed9/10

   GrowthBook's quickstart docs show flag creation in a few lines of code with SDK hooks like useFeatureIsOn, and a runtime probe confirms a real keyless create-and-evaluate loop (create org, create flag, mint SDK connection, get correct value back) with no signup friction. Independent community testimony corroborates fast setup, with one user reporting a working flag in about 15 minutes and under 10 lines of code, and environments docs confirm dev/staging/prod toggling. missing for 10: a first-party or independent account of toggling a flag live specifically in a production environment end-to-end.

   • [claimed-docs] “Get GrowthBook feature flags and experiments running in your app in a few lines of code, then pick a language-specific SDK.”
   • [claimed-docs] “const showNewFeature = useFeatureIsOn("new-feature");”
   • [claimed-docs] “Use environments such as dev, staging, and production to enable or disable feature flags independently.”
   • [community] “Been using GrowthBook for a few weeks now... I was shocked at how much setup time it took to get a simple feature flag into our app with oth…”
   • [probe] “PROBE runtime (recorded 2026-09-05): full KEYLESS self-host roundtrip on this machine — the official growthbook/growthbook docker image boot…”

   proof ↗⚑ flag

2. developerServe multivariate flags and dynamic configuration values (strings, numbers, JSON), not just booleans

   weight 2 · round to GrowthBook
   LaunchDarkly~partialcommunity6/10

   Community evidence explicitly confirms multivariate flags as a key LaunchDarkly benefit (launchdarkly-comm-9), and docs reference custom rules, targeting, and percentage rollouts that imply non-boolean variation support, but no evidence pack item directly documents string/number/JSON typed flag variations or dynamic config value usage. missing for 10: explicit docs on flag variation types (string/number/JSON), examples of dynamic config values, independent hands-on confirmation beyond the single HN mention.

   • [community] “Key benefits of LaunchDarkly: rules-based targeting so flags can be set at group level, custom rollout by % of users or attributes, centrali…”
   • [claimed-docs] “Targeting rules”
   • [claimed-docs] “Percentage rollouts”
   • [claimed-docs] “Custom rules”

   proof ↗⚑ flag

   GrowthBook✓fullprobed7/10

   GrowthBook's SDK docs explicitly show non-boolean flag values via useFeatureValue (string example 'button-color') alongside useFeatureIsOn for booleans, and its targeting/rules docs describe rules that 'control which values a user sees' beyond simple on/off, plus a real runtime probe confirms a flag payload with a typed defaultValue served via the API. Missing for 10: explicit documented example of numeric and JSON-typed flag values (only string is shown in the quickstart snippet), and no independent hands-on confirmation of JSON config usage.

   • [claimed-docs] “const showNewFeature = useFeatureIsOn("new-feature"); const buttonColor = useFeatureValue("button-color", "blue");”
   • [claimed-docs] “How rules override a feature's default value — targeting users, rolling out gradually, experimenting — and how to scope, ramp, and schedule …”
   • [claimed-docs] “Add targeting conditions to a rule to control which values a user sees.”
   • [probe] “PROBE runtime (recorded 2026-09-05): full KEYLESS self-host roundtrip on this machine — the official growthbook/growthbook docker image boot…”
   • [github] “24 SDKs including React, Python, Android, and iOS.”

   proof ↗⚑ flag

Lifecycle

1. platform engineerFind stale flags and code references so temporary flags actually get removed from the codebase

   weight 2 · round to GrowthBook
   LaunchDarkly~partialcommunity6/10

   LaunchDarkly documents a dedicated 'Code References' feature that shows where flags are used in code, corroborated by independent community commentary ([launchdarkly-docs-22], [launchdarkly-comm-3]), which directly supports finding code references for stale flags. However, the evidence pack lacks explicit documentation of flag-status/stale-flag detection (e.g., 'last evaluated' tracking or automated stale-flag alerts) to complete the workflow. Missing for 10: dedicated stale-flag status/health dashboard docs, evidence of automated staleness alerts, and independent confirmation that code-reference removal workflows are reliably used to clean up flags in practice.

   • [claimed-docs] “Code references”
   • [community] “One of the most common things you want to do for a feature flag is ask 'Where is this used in code?' LaunchDarkly even has a product feature…”

   proof ↗⚑ flag

   GrowthBook✓fullclaimed8/10

   GrowthBook's docs explicitly describe automatic stale flag detection ('GrowthBook can automatically detect these flags so you can clean them up') as a dedicated feature, directly addressing the story of finding stale flags for removal. This is corroborated by the product's broader flag-management infrastructure (environments, rules, revisions) but lacks detail on code-reference scanning across repos or independent/hands-on verification of the stale-detection feature itself. Missing for 10: evidence of code-reference/usage scanning across the codebase, and independent/community confirmation that stale detection works reliably in practice.

   • [claimed-docs] “GrowthBook can automatically detect these flags so you can clean them up.”
   • [claimed-docs] “How rules override a feature's default value — targeting users, rolling out gradually, experimenting — and how to scope, ramp, and schedule …”
   • [claimed-docs] “Use environments such as dev, staging, and production to enable or disable feature flags independently.”

   proof ↗⚑ flag

Targeting

1. developerTarget or exclude specific individual users for a flag (allowlists, beta testers, internal accounts)

   weight 1 · round to LaunchDarkly
   LaunchDarkly✓fullcommunity8/10

   LaunchDarkly's docs explicitly cover 'Individual targeting' and 'Target with flags' allowing developers to include/exclude specific users by key, plus segments and custom rules for building allowlists (beta testers, internal accounts), corroborated by community mention of rules-based group-level targeting. Missing for 10: independent hands-on verification of individual-user allowlist workflow, dedicated example of internal-account exclusion in practice.

   • [claimed-docs] “Targeting rules”
   • [claimed-docs] “Individual targeting”
   • [claimed-docs] “Target with flags”
   • [claimed-docs] “Segments”
   • [claimed-docs] “Custom rules”
   • [community] “Key benefits of LaunchDarkly: rules-based targeting so flags can be set at group level, custom rollout by % of users or attributes, centrali…”

   proof ↗⚑ flag

   GrowthBook✓fullclaimed7/10

   GrowthBook's rule/targeting system lets you write conditions against arbitrary user attributes (e.g., user ID, email) to include or exclude specific users, which covers allowlisting beta testers or internal accounts (growthbook-docs-2, growthbook-docs-22, growthbook-docs-18, growthbook-docs-28). This is a first-class part of its rules engine (scope, ramp, schedule). missing for 10: an explicit doc example/tutorial titled 'allowlist' or 'beta testers' walkthrough, and independent hands-on confirmation of exclusion-list behavior.

   • [claimed-docs] “Add targeting conditions to a rule to control which values a user sees. Conditions are evaluated against attributes you pass into the SDK”
   • [claimed-docs] “How rules override a feature's default value — targeting users, rolling out gradually, experimenting — and how to scope, ramp, and schedule …”
   • [claimed-docs] “The actual values of targeting attributes (e.g., user IDs, emails) are never sent to GrowthBook. They are only stored in memory locally with…”
   • [claimed-docs] “Add targeting conditions to a rule to control which values a user sees.”

   proof ↗⚑ flag

2. developerTarget flags with attribute-based rules and reusable segments so the right users see the right variation

   weight 3 · round to LaunchDarkly
   LaunchDarkly✓fullcommunity9/10

   LaunchDarkly's docs and community evidence confirm attribute-based targeting rules, contexts, individual targeting, custom rules, percentage rollouts, and reusable segments (including converting rules into segments), directly matching the story; community input corroborates rules-based targeting by group/attributes in production use. missing for 10: independent hands-on verification of segment reuse across multiple flags beyond docs claims.

   • [claimed-docs] “Targeting rules”
   • [claimed-docs] “Individual targeting”
   • [claimed-docs] “Contexts”
   • [claimed-docs] “Segments”
   • [claimed-docs] “Converting rules into segments”
   • [claimed-docs] “Custom rules”
   • [claimed-docs] “Target with flags”
   • [community] “Key benefits of LaunchDarkly: rules-based targeting so flags can be set at group level, custom rollout by % of users or attributes, centrali…”

   proof ↗⚑ flag

   GrowthBook~partialclaimed6/10

   GrowthBook clearly supports attribute-based targeting rules (conditions evaluated against SDK-passed attributes, rule scoping/ramping/scheduling) as shown in growthbook-docs-2, -22, -28, and -18. However, the evidence pack never mentions a distinct 'segments' feature (reusable named audience definitions) that developers could apply across multiple flags/rules, so the reusable-segment half of the story is unevidenced. missing for 10: explicit segment/audience-reuse documentation, evidence of applying the same segment across multiple flags, independent confirmation of segment behavior.

   • [claimed-docs] “Add targeting conditions to a rule to control which values a user sees. Conditions are evaluated against attributes you pass into the SDK”
   • [claimed-docs] “How rules override a feature's default value — targeting users, rolling out gradually, experimenting — and how to scope, ramp, and schedule …”
   • [claimed-docs] “Add targeting conditions to a rule to control which values a user sees.”
   • [claimed-docs] “The actual values of targeting attributes (e.g., user IDs, emails) are never sent to GrowthBook. They are only stored in memory locally with…”

   proof ↗⚑ flag

Access

1. platform engineerRestrict who can change which flags with roles, permissions, and scoped API tokens

   weight 2 · round to LaunchDarkly
   LaunchDarkly✓fullclaimed8/10

   LaunchDarkly documents role-based access control ('Assigning roles to members') and scoped API access tokens ('API access tokens'), directly matching the story's request for roles, permissions, and scoped tokens to restrict who can change flags. Missing for 10: independent/hands-on verification of custom role granularity and confirmation of per-token flag/project scoping restrictions in practice, and no community corroboration of this specific governance feature.

   • [claimed-docs] “Assigning roles to members”
   • [claimed-docs] “API access tokens”

   proof ↗⚑ flag

   GrowthBook~partialclaimed5/10

   Pricing page confirms GrowthBook has role/permission features ("Advanced permissioning", "Advanced access control", SSO & SCIM) gated behind paid tiers, implying RBAC exists, but there is no docs page detailing how roles map to flag-level permissions or how API tokens can be scoped to restrict flag changes; the API docs even note the SDK connection endpoint is public/unauthenticated, which is a different, more limited access model. Missing for 10: a dedicated permissions/roles doc, explicit description of scoped/restricted API tokens for flag mutation, and independent confirmation of enforcement.

   • [claimed-docs] “Starter, plus: Visual editor Multi-arm bandits Safe rollouts Customizable dashboards Advanced permissioning Power calculator Premium support”
   • [claimed-docs] “Pro, plus: Custom environments Ramp schedules Approval workflows Advanced access control SSO & SCIM provisioning Exportable audit logs”
   • [claimed-docs] “Visual editor * Multi-arm bandits * Safe rollouts * Customizable dashboards * Advanced permissioning * Power calculator”
   • [claimed-docs] “Custom environments * Ramp schedules * Approval workflows * Advanced access control * SSO & SCIM provisioning * Exportable audit l…”
   • [claimed-docs] “GrowthBook offers a full REST API for interacting with the application.”
   • [claimed-docs] “They are meant to be public and do not require authentication to view.”

   proof ↗⚑ flag

Agent ops

1. ai agentCreate and toggle flags through documented APIs, CLIs, or MCP — and the platform can force my changes through approval workflows instead of letting me write to production unreviewed

   weight 3 · round to LaunchDarkly
   LaunchDarkly✓fullprobed7/10

   LaunchDarkly documents an official CLI (ldcli) and both local and hosted MCP servers for programmatic flag creation/toggling, backed by API access tokens and RBAC roles, and runtime probes confirm the CLI and MCP handshakes actually work as documented. Separately, LaunchDarkly's Approvals feature lets teams require review/sign-off before flag changes go live, and Change History provides an audit trail, which together satisfy the 'force changes through approval workflows' requirement. Missing for 10: explicit documentation tying the Approvals workflow specifically to API/CLI/MCP-originated changes (vs. UI-only), and independent/hands-on confirmation that an agent's API write is actually blocked pending approval.

   • [claimed-docs] “Approvals”
   • [claimed-docs] “Change history”
   • [claimed-docs] “LaunchDarkly MCP server”
   • [claimed-docs] “Set up the LaunchDarkly hosted MCP server”
   • [claimed-docs] “LaunchDarkly CLI”
   • [claimed-docs] “Assigning roles to members”
   • [claimed-docs] “API access tokens”
   • [probe] “official MCP server documented at https://launchdarkly.com/docs/home/getting-started/mcp”
   • [probe] “official CLI documented at https://launchdarkly.com/docs/home/getting-started/ldcli”
   • [probe] “PROBE runtime (recorded 2026-09-05): the official LaunchDarkly CLI installed via `brew tap launchdarkly/homebrew-tap && brew install ldcli` …”
   • [probe] “PROBE runtime (recorded 2026-09-05): the official npm MCP server (`npx -y @launchdarkly/mcp-server start --transport stdio`) completed a FUL…”
   • [probe] “PROBE runtime (recorded 2026-09-05): keyless JSON-RPC initialize POST to the documented hosted MCP endpoint https://mcp.launchdarkly.com/mcp…”

   proof ↗⚑ flag

   GrowthBook~partialprobed6/10

   GrowthBook documents an official MCP server that makes authenticated REST calls (growthbook-docs-13/21, growthbook-probe-3/rt-1) and a full REST API (growthbook-docs-12), and separately documents 'Draft feature revisions, request review... require approval before publishing' as a Pro+ approval-workflow feature (growthbook-docs-5, growthbook-docs-16/37). However, the runtime probe (growthbook-probe-rt-2) shows a flag created and immediately live via the API with no review step, and approval workflows appear gated behind a paid plan rather than proven to force-block an AI agent's writes by default. Missing for 10: a documented/verified CLI channel, and hands-on evidence that MCP/API writes are actually intercepted and blocked pending human approval rather than merely available as an opt-in paid feature.

   • [claimed-docs] “Connect AI tools like Cursor, VS Code, and Claude to GrowthBook using the Model Context Protocol (MCP).”
   • [claimed-docs] “The official GrowthBook MCP server is a thin bridge: it loads GrowthBook Agent Skills (workflows and guardrails) and makes authenticated cal…”
   • [probe] “official MCP server documented at https://docs.growthbook.io/integrations/mcp”
   • [probe] “PROBE runtime (recorded 2026-09-05): the official npm MCP server (`npx -y @growthbook/mcp`) completed a FULL keyless stdio initialize handsh…”
   • [claimed-docs] “Draft feature revisions, request review, resolve merge conflicts, and require approval before publishing.”
   • [claimed-docs] “Pro, plus: Custom environments Ramp schedules Approval workflows Advanced access control SSO & SCIM provisioning Exportable audit logs”
   • [claimed-docs] “Custom environments * Ramp schedules * Approval workflows * Advanced access control * SSO & SCIM provisioning * Exportable audit l…”
   • [claimed-docs] “GrowthBook offers a full REST API for interacting with the application.”
   • [probe] “PROBE runtime (recorded 2026-09-05): full KEYLESS self-host roundtrip on this machine — the official growthbook/growthbook docker image boot…”

   proof ↗⚑ flag

Approvals

1. platform engineerRequire approvals or change requests before production flag changes go live

   weight 3 · round drawn
   LaunchDarkly✓fullclaimed8/10

   LaunchDarkly documents a dedicated Approvals feature requiring sign-off before flag changes are applied, plus Scheduled changes and Change history for governance/audit trails around production flag modifications, backed by role-based access control to enforce who can approve. missing for 10: independent hands-on confirmation of the approval workflow in practice and details on customizable approval policies (e.g., required approver count, bypass rules).

   • [claimed-docs] “Approvals”
   • [claimed-docs] “Change history”
   • [claimed-docs] “Scheduled flag changes”
   • [claimed-docs] “Assigning roles to members”

   proof ↗⚑ flag

   GrowthBook✓fullclaimed8/10

   GrowthBook's docs explicitly describe drafting feature revisions, requesting review, resolving merge conflicts, and requiring approval before publishing changes (growthbook-docs-5), and this is packaged as a named 'Approval workflows' feature on paid tiers (growthbook-docs-16, growthbook-docs-37). This directly matches the story of gating production flag changes behind approvals. Missing for 10: independent/hands-on confirmation of the approval workflow in practice (only vendor docs cited) and detail on how approvals interact with environment-specific promotion.

   • [claimed-docs] “Draft feature revisions, request review, resolve merge conflicts, and require approval before publishing.”
   • [claimed-docs] “Pro, plus: Custom environments Ramp schedules Approval workflows Advanced access control SSO & SCIM provisioning Exportable audit logs”
   • [claimed-docs] “Custom environments * Ramp schedules * Approval workflows * Advanced access control * SSO & SCIM provisioning * Exportable audit l…”
   • [claimed-docs] “Use environments such as dev, staging, and production to enable or disable feature flags independently.”

   proof ↗⚑ flag

Audit

1. platform engineerEvery flag change is recorded in an audit log — who changed what, when, and to which value

   weight 3 · round to LaunchDarkly
   LaunchDarkly~partialcommunity6/10

   LaunchDarkly documents 'Change history' for flags, which tracks flag configuration changes, and community mentions of Slack integration for visibility into changes corroborates change tracking exists; however, no evidence explicitly confirms a full audit log capturing who/when/old-value/new-value in a governance-focused, exportable/immutable format. Missing for 10: explicit documentation of audit-log fields (actor, timestamp, before/after value), retention/export/SIEM integration details, and independent confirmation of completeness.

   • [claimed-docs] “Change history”
   • [community] “Key benefits of LaunchDarkly: rules-based targeting so flags can be set at group level, custom rollout by % of users or attributes, centrali…”

   proof ↗⚑ flag

   GrowthBook~partialclaimed5/10

   Pricing docs confirm GrowthBook ships an audit log feature, with 'Exportable audit logs' listed as a Pro-tier capability, implying built-in tracking of changes for governance purposes. However, there is no first-party documentation detailing exactly what is captured (who/when/old-new value) or how the audit trail is presented/queried, and the capability is gated behind a paid plan rather than universally available. Missing for 10: dedicated audit-log documentation describing captured fields, retention, and access/export mechanics, plus confirmation it's available outside paid tiers or independent corroboration of its behavior.

   • [claimed-docs] “Pro, plus: Custom environments Ramp schedules Approval workflows Advanced access control SSO & SCIM provisioning Exportable audit logs”
   • [claimed-docs] “Custom environments * Ramp schedules * Approval workflows * Advanced access control * SSO & SCIM provisioning * Exportable audit l…”
   • [claimed-docs] “SSO & SCIM provisioning * Exportable audit logs”

   proof ↗⚑ flag

1. ai-native userDo everything through the API that I can do in the UI

   weight 2 · round to GrowthBook
   LaunchDarkly~partialprobed5/10

   LaunchDarkly documents API access tokens for programmatic access, an official CLI (ldcli, verified installable and working) and both a local and hosted MCP server (verified live and functioning), indicating broad programmatic control over flags, targeting, and other resources. However, the evidence pack never explicitly states or demonstrates full parity between UI and API (e.g., no OpenAPI/swagger spec was discoverable, and no docs confirm every UI feature—like approvals, guarded rollouts, or experimentation setup—is API-exposed). Missing for 10: an explicit statement or reference doc confirming 100% UI/API feature parity, and evidence covering API support for approvals, guarded rollouts, and experimentation configuration specifically.

   • [claimed-docs] “API access tokens”
   • [claimed-docs] “LaunchDarkly CLI”
   • [probe] “official MCP server documented at https://launchdarkly.com/docs/home/getting-started/mcp”
   • [probe] “official CLI documented at https://launchdarkly.com/docs/home/getting-started/ldcli”
   • [probe] “PROBE runtime (recorded 2026-09-05): the official LaunchDarkly CLI installed via `brew tap launchdarkly/homebrew-tap && brew install ldcli` …”
   • [probe] “PROBE runtime (recorded 2026-09-05): the official npm MCP server (`npx -y @launchdarkly/mcp-server start --transport stdio`) completed a FUL…”
   • [probe] “PROBE runtime (recorded 2026-09-05): keyless JSON-RPC initialize POST to the documented hosted MCP endpoint https://mcp.launchdarkly.com/mcp…”
   • [probe] “PROBE openapi: all candidate paths 404 (https://launchdarkly.com/openapi.json, https://launchdarkly.com/swagger.json, https://launchdarkly.c…”

   proof ↗⚑ flag

   GrowthBook✓fullprobed7/10

   GrowthBook documents a 'full REST API for interacting with the application' (growthbook-docs-12) and a runtime probe confirms it: a self-hosted instance created an org, user, and feature flag purely via API calls and served it through the SDK payload endpoint (growthbook-probe-rt-2), and the official MCP server exposes authenticated read/write meta-tools covering the whole REST API (growthbook-docs-21, growthbook-probe-rt-1). Missing for 10: an explicit enumerated API-vs-UI parity list (e.g. confirmation that experiment approval workflows, visual editor changes, and dashboard configuration are all API-exposed, not just feature flags) and independent third-party corroboration of API completeness.

   • [claimed-docs] “GrowthBook offers a full REST API for interacting with the application.”
   • [claimed-docs] “The official GrowthBook MCP server is a thin bridge: it loads GrowthBook Agent Skills (workflows and guardrails) and makes authenticated cal…”
   • [probe] “PROBE runtime (recorded 2026-09-05): the official npm MCP server (`npx -y @growthbook/mcp`) completed a FULL keyless stdio initialize handsh…”
   • [probe] “PROBE runtime (recorded 2026-09-05): full KEYLESS self-host roundtrip on this machine — the official growthbook/growthbook docker image boot…”
   • [claimed-docs] “The SDK Connection Endpoint provides readonly access to a subset of your feature flag data, just enough for the GrowthBook SDKs to assign va…”

   proof ↗⚑ flag

2. ai-native userExport all of my data in open formats and leave

   weight 3 · round to GrowthBook
   LaunchDarkly—none0/10

   LaunchDarkly offers API access tokens and a CLI (ldcli) that could programmatically read flag/segment configurations, but no evidence pack item documents a dedicated bulk data-export feature, an open-format export tool, or a stated data-portability/exit policy for account data.

   • [claimed-docs] “API access tokens”
   • [claimed-docs] “LaunchDarkly CLI”

   proof ↗⚑ flag

   GrowthBook~partialcommunity5/10

   GrowthBook is open-source and self-hostable with its data stored in a standard MongoDB instance, and it exposes a full REST API for programmatic access to feature flags, experiments, and metrics, which together give a technically-inclined user a path to extract all their data without vendor lock-in. However, there is no explicit documented 'export all data' or bulk-export feature/wizard, and 'exportable audit logs' is gated behind an Enterprise tier rather than being a general open-export capability. Missing for 10: an explicit documented full-data export/backup tool or format, confirmation that experiment results/metrics history export in an open format, and evidence this works outside self-hosting for cloud customers.

   • [claimed-docs] “GrowthBook offers a full REST API for interacting with the application.”
   • [claimed-docs] “Run GrowthBook on your own infrastructure with Docker Compose, including MongoDB, the API, and the front-end.”
   • [claimed-docs] “You can use docker compose to get started quickly... run docker compose up -d to start everything and view the app at http://localhost:3000”
   • [claimed-docs] “Custom environments * Ramp schedules * Approval workflows * Advanced access control * SSO & SCIM provisioning * Exportable audit l…”
   • [community] “Re: Eppo - they're similar in setup but Growthbook is open source with an optional cloud SaaS offering. Optimizely is a walled garden, you n…”

   proof ↗⚑ flag

3. ai-native userRead the product's source under an open license

   weight 2 · round to GrowthBook
   LaunchDarkly—none0/10

   LaunchDarkly is a closed-source SaaS platform; evidence shows only proprietary docs, SDKs (some open-source client SDKs and CLI/relay proxy tools exist per community notes, but the core product source is not open-licensed), with no indication the core platform source is available under an open license.

   • [community] “LaunchDarkly clients download all flags in the project on connection, then stream changes. ... LaunchDarkly has an open source 'Relay' you c…”

   proof ↗⚑ flag

   GrowthBook~partialcommunity5/10

   Community evidence explicitly states GrowthBook is open source with an optional cloud SaaS offering, and a public GitHub repository (github.com/growthbook/growthbook) is cited, supporting that source is readable. However, no evidence pack item specifies the actual license type (e.g., MIT/Apache) or licensing terms, so the 'open license' claim isn't directly documented. Missing for 10: explicit license identification/text, first-party docs confirming license terms, independent confirmation of license permissiveness.

   • [community] “Re: Eppo - they're similar in setup but Growthbook is open source with an optional cloud SaaS offering. Optimizely is a walled garden, you n…”
   • [github] “24 SDKs including React, Python, Android, and iOS.”
   • [claimed-docs] “Run GrowthBook on your own infrastructure with Docker Compose, including MongoDB, the API, and the front-end.”

   proof ↗⚑ flag

4. ai-native userSelf-host the core product

   weight 3 · round to GrowthBook
   LaunchDarkly—none0/10

   LaunchDarkly is a hosted SaaS platform; the only self-hostable component is the Relay Proxy, a caching/streaming layer that still depends on the LaunchDarkly SaaS control plane for flag configuration, not the core product itself. No evidence of a self-hostable core (dashboard, rules engine, data store) exists in the pack.

   • [claimed-docs] “The LaunchDarkly Relay Proxy is a small Go application that runs on your own infrastructure.”
   • [claimed-docs] “The Relay Proxy”
   • [community] “LaunchDarkly clients download all flags in the project on connection, then stream changes. ... LaunchDarkly has an open source 'Relay' you c…”
   • [community] “It's a pretty question[able] architecture that outsources feature flags to a third party. You're creating a critical path hard dependency an…”

   proof ↗⚑ flag

   GrowthBook✓fullprobed10/10

   GrowthBook is documented and independently confirmed (via runtime probe) to self-host fully via Docker Compose including MongoDB, API, and front-end, with a real keyless create-and-evaluate flag loop working end-to-end; community reviews also corroborate long-term self-hosted usage.

   • [claimed-docs] “Run GrowthBook on your own infrastructure with Docker Compose, including MongoDB, the API, and the front-end.”
   • [claimed-docs] “You can use docker compose to get started quickly... run docker compose up -d to start everything and view the app at http://localhost:3000”
   • [probe] “PROBE runtime (recorded 2026-09-05): full KEYLESS self-host roundtrip on this machine — the official growthbook/growthbook docker image boot…”
   • [community] “Growthbook is awesome! We have used the self-hosted version a few months in my company now, and are very happy about it. Talented people beh…”

   proof ↗⚑ flag

Pricing

1. product managerSee published pricing and understand what drives cost (seats, MAUs, events, requests) before committing

   weight 2 · round drawn
   LaunchDarkly~partialcommunity6/10

   LaunchDarkly's public pricing page confirms a usage-based model (unlimited seats, no per-seat pricing) and lists some usage components like AI Runs and session-replay/error caps, plus community confirmation that mobile MAUs factor into cost. However, the evidence pack lacks concrete published tiers/numbers for events or API requests, and community commentary calls the pricing 'incredibly over priced' and confusing to reason about ahead of committing. missing for 10: detailed published rate tables for MAUs/events/requests, a cost calculator or worked pricing examples, and independent confirmation that the pricing page is fully transparent pre-sales.

   • [claimed-docs] “Configure, evaluate, roll out, and observe your AI end to end”
   • [claimed-docs] “An AI Run is generated when you track an LLM request or run an online eval (LLM-as-a-judge).”
   • [claimed-docs] “Unlimited feature flags”
   • [claimed-docs] “5K session replays and 5K errors /mo”
   • [claimed-docs] “Unlimited seats No per-seat pricing. Invite your whole team; pricing is based on usage.”
   • [community] “LaunchDarkly has a mobile client SDK, but they charge by MAU, which would be untenable. So folks tend to write a proxy for the mobile apps t…”
   • [community] “Agreed that LaunchDarkly is incredibly over priced.”

   proof ↗⚑ flag

   GrowthBook~partialclaimed6/10

   GrowthBook publishes a tiered pricing page (Starter/Pro/Enterprise-like tiers) listing feature gates (visual editor, approval workflows, SSO/SCIM, etc.) and calls out unlimited flags/experiments/traffic and per-user caps like 'Up to 3 users', but none of the evidence shows usage-based cost drivers like MAUs, events, or API requests being priced or metered — the plans appear seat/feature-gated rather than usage-metered. missing for 10: explicit cost drivers tied to MAUs/events/requests, a pricing calculator or usage-based cost breakdown, and independent confirmation that published prices match real invoices.

   • [claimed-docs] “Starter, plus: Visual editor Multi-arm bandits Safe rollouts Customizable dashboards Advanced permissioning Power calculator Premium support”
   • [claimed-docs] “Pro, plus: Custom environments Ramp schedules Approval workflows Advanced access control SSO & SCIM provisioning Exportable audit logs”
   • [claimed-docs] “Unlimited feature flags * Unlimited experiments * Unlimited traffic”
   • [claimed-docs] “Starter, plus: * Visual editor * Multi-arm bandits * Safe rollouts * Customizable dashboards”
   • [claimed-docs] “Visual editor * Multi-arm bandits * Safe rollouts * Customizable dashboards * Advanced permissioning * Power calculator”
   • [claimed-docs] “Custom environments * Ramp schedules * Approval workflows * Advanced access control * SSO & SCIM provisioning * Exportable audit l…”
   • [claimed-docs] “Up to 3 users * 1 project * Unlimited feature flags * Unlimited experiments * Unlimited traffic”
   • [claimed-docs] “SSO & SCIM provisioning * Exportable audit logs”

   proof ↗⚑ flag

1. ai-native userChoose where my data is stored (region/residency)

   weight 2 · round to GrowthBook
   LaunchDarkly—none0/10

   No evidence pack items address data residency, region selection, or storage location controls for LaunchDarkly; the only tangential item is 'federal environments' docs, which is not about general data residency choice. missing for 10: documentation of selectable data regions/residency, EU/US hosting options, or any data localization controls.

   • [claimed-docs] “LaunchDarkly in federal environments”

   proof ↗⚑ flag

   GrowthBook~partialprobed6/10

   GrowthBook can be fully self-hosted via Docker Compose (own MongoDB, API, and front-end), which lets an AI-native user control exactly where all data lives, and a probe confirms a working self-hosted flag create/evaluate loop with no third-party data storage. However, there is no documented cloud-hosted region/residency selection (e.g., EU vs US SaaS options) for users who don't want to self-host. Missing for 10: explicit cloud multi-region hosting options, formal data-residency/compliance certifications, and documentation letting SaaS customers pick a storage region without self-hosting.

   • [claimed-docs] “Run GrowthBook on your own infrastructure with Docker Compose, including MongoDB, the API, and the front-end.”
   • [claimed-docs] “You can use docker compose to get started quickly... run docker compose up -d to start everything and view the app at http://localhost:3000”
   • [probe] “PROBE runtime (recorded 2026-09-05): full KEYLESS self-host roundtrip on this machine — the official growthbook/growthbook docker image boot…”
   • [claimed-docs] “The actual values of targeting attributes (e.g., user IDs, emails) are never sent to GrowthBook. They are only stored in memory locally with…”

   proof ↗⚑ flag

2. ai-native userControl data retention and deletion

   weight 2 · round to GrowthBook
   LaunchDarkly—none0/10

   LaunchDarkly is a feature-flag/experimentation platform; while it stores user/context data (contexts, segments), no evidence in the pack addresses data retention policies, deletion controls, or privacy/compliance tooling for AI-native users. missing for 10: data retention policy docs, deletion/erasure API or workflow, GDPR/CCPA data-subject-request handling, retention configuration options.

   ⚑ flag

   GrowthBook~partialclaimed4/10

   GrowthBook's privacy-by-design targeting (attribute values never sent to GrowthBook, kept in-memory in the SDK) and the option to self-host on your own infrastructure/MongoDB give users indirect control over what data is retained and how it's deleted, since they own the database. However, there is no explicit documentation of a data retention policy, a deletion API/endpoint, or GDPR-style 'delete my data' tooling in the evidence pack. Missing for 10: explicit retention-period settings, a documented deletion/erasure API or workflow, and any compliance certification (e.g., GDPR/CCPA) statements.

   • [claimed-docs] “The actual values of targeting attributes (e.g., user IDs, emails) are never sent to GrowthBook. They are only stored in memory locally with…”
   • [claimed-docs] “Run GrowthBook on your own infrastructure with Docker Compose, including MongoDB, the API, and the front-end.”
   • [claimed-docs] “You can use docker compose to get started quickly... run docker compose up -d to start everything and view the app at http://localhost:3000”

   proof ↗⚑ flag

3. ai-native userOpt out of telemetry and usage tracking

   weight 2 · round drawn
   LaunchDarkly—none0/10

   No evidence pack item addresses telemetry/usage-tracking opt-out for LaunchDarkly itself; the docs cover flagging, targeting, MCP/CLI tooling, and pricing but nothing about the product's own data collection practices or an opt-out mechanism.

   ⚑ flag

   GrowthBook—none0/10

   The evidence pack shows GrowthBook is self-hostable and that targeting-attribute values are never sent to the vendor, but there is no documentation of any product telemetry/usage-tracking opt-out setting or flag for the self-hosted or cloud application itself. Missing for 10: any explicit telemetry/usage-analytics toggle, docs on what data is phoned home from self-hosted instances, or a documented opt-out mechanism.

   ⚑ flag

Rollouts

1. platform engineerGuard a rollout with metrics so a regression is detected and the release is rolled back automatically

   weight 2 · round to LaunchDarkly
   LaunchDarkly~partialclaimed6/10

   LaunchDarkly's Guarded Rollouts feature explicitly monitors metrics during a progressive rollout and automatically rolls back on regression detection, directly matching the story, and is backed by supporting docs on metrics, progressive rollouts, and change history. However, the evidence pack is entirely first-party docs with no independent/hands-on confirmation that automatic rollback actually triggers correctly in practice, and no detail on metric types supported for regression detection or configuration limits. Missing for 10: independent/hands-on validation of guarded rollout auto-rollback behavior, detail on which metric types/thresholds trigger rollback, and confirmation of reliability at scale.

   • [claimed-docs] “Guarded rollouts”
   • [claimed-docs] “Creating and managing progressive rollouts”
   • [claimed-docs] “Choosing a metric type”
   • [claimed-docs] “Change history”

   proof ↗⚑ flag

   GrowthBook~partialclaimed4/10

   GrowthBook's pricing pages list a named 'Safe rollouts' feature (Pro+ tier) alongside guardrail-metric support in experiment templates, implying automated regression detection and rollback, but the evidence pack contains no docs explaining how the automatic rollback mechanism actually works (thresholds, trigger logic, rollback speed) nor any hands-on confirmation it fires reliably in production. Missing for 10: detailed docs/description of the Safe Rollouts mechanism and guardrail-triggered auto-rollback, plus independent or hands-on evidence it works as claimed.

   • [claimed-docs] “Starter, plus: Visual editor Multi-arm bandits Safe rollouts Customizable dashboards Advanced permissioning Power calculator Premium support”
   • [claimed-docs] “Starter, plus: * Visual editor * Multi-arm bandits * Safe rollouts * Customizable dashboards”
   • [claimed-docs] “Visual editor * Multi-arm bandits * Safe rollouts * Customizable dashboards * Advanced permissioning * Power calculator”
   • [claimed-docs] “Visual editor * Multi-arm bandits * Safe rollouts * Customizable dashboards”
   • [claimed-docs] “Selecting a Template in the creation modal pre-fills settings like the data source, goal/secondary/guardrail metrics, targeting, and traffic”

   proof ↗⚑ flag

2. developerRoll a flag out progressively by percentage with consistent bucketing, ramping from 1% to 100% without redeploying

   weight 3 · round to LaunchDarkly
   LaunchDarkly✓fullcommunity9/10

   LaunchDarkly's core product is percentage-based progressive rollouts with consistent bucketing via context attributes, explicitly documented (progressive rollouts, percentage rollouts, guarded rollouts) and corroborated by community users citing 'custom rollout by % of users or attributes' as a key benefit, all changed instantly without redeploy through the flag dashboard/API. Missing for 10: no independent hands-on verification of bucketing consistency across ramp changes or a documented technical explanation of the hashing algorithm.

   • [claimed-docs] “Creating and managing progressive rollouts”
   • [claimed-docs] “Guarded rollouts”
   • [claimed-docs] “Percentage rollouts”
   • [community] “Key benefits of LaunchDarkly: rules-based targeting so flags can be set at group level, custom rollout by % of users or attributes, centrali…”

   proof ↗⚑ flag

   GrowthBook✓fullprobed8/10

   GrowthBook's rules docs explicitly cover 'rolling out gradually' and describe how rules can 'scope, ramp, and schedule' percentage rollouts, with experiment/flag assignment based on a hashing attribute for consistent bucketing, and SDKs evaluate the cached JSON payload client-side with no redeploy needed. Pricing tiers also list dedicated 'Ramp schedules' as a named capability, confirming this is a first-class shipped feature, and probe evidence confirms a live flag create/evaluate loop via the SDK payload endpoint. missing for 10: independent/hands-on confirmation of an actual 1%→100% ramp in practice, and detail on the specific hashing algorithm ensuring bucket stability across ramp changes

   • [claimed-docs] “How rules override a feature's default value — targeting users, rolling out gradually, experimenting — and how to scope, ramp, and schedule …”
   • [claimed-docs] “you can add an experiment rule to a feature that will randomly assign the users based on some hashing attribute into one of your experiment …”
   • [claimed-docs] “When the client loads your application, our SDK will check to see if the user should be part of any experiments, and if so, assign them and …”
   • [claimed-docs] “Pro, plus: Custom environments Ramp schedules Approval workflows Advanced access control SSO & SCIM provisioning Exportable audit logs”
   • [claimed-docs] “Custom environments * Ramp schedules * Approval workflows * Advanced access control * SSO & SCIM provisioning * Exportable audit l…”
   • [claimed-docs] “GrowthBook supports several ways to do this: Feature Flags (most common)... URL Redirects... Our Visual Editor... Your own custom variation …”
   • [probe] “PROBE runtime (recorded 2026-09-05): full KEYLESS self-host roundtrip on this machine — the official growthbook/growthbook docker image boot…”

   proof ↗⚑ flag

Scheduling

1. developerSchedule flag changes and releases to happen at a specific future time

   weight 1 · round to LaunchDarkly
   LaunchDarkly✓fullclaimed9/10

   LaunchDarkly documents a dedicated 'Scheduled flag changes' feature allowing flag changes to be set to take effect at a future time, alongside related release tooling like progressive/guarded rollouts and change history for tracking scheduled edits. Missing for 10: independent hands-on confirmation or community corroboration of the scheduling workflow specifically (only vendor docs cited).

   • [claimed-docs] “Scheduled flag changes”
   • [claimed-docs] “Creating and managing progressive rollouts”
   • [claimed-docs] “Guarded rollouts”
   • [claimed-docs] “Change history”

   proof ↗⚑ flag

   GrowthBook~partialclaimed6/10

   GrowthBook explicitly documents that feature rules can be 'scoped, ramped, and scheduled,' and its pricing page lists 'Ramp schedules' as a distinct capability, confirming a future-time scheduling mechanism for flag rollouts. However, this appears gated to paid plans (Pro tier) and the evidence pack gives no detailed walkthrough of how scheduling is configured or any independent/hands-on confirmation it works as described. Missing for 10: detailed documentation of the scheduling UI/API, hands-on or community verification, and clarity on whether it applies broadly to all rule types.

   • [claimed-docs] “How rules override a feature's default value — targeting users, rolling out gradually, experimenting — and how to scope, ramp, and schedule …”
   • [claimed-docs] “Pro, plus: Custom environments Ramp schedules Approval workflows Advanced access control SSO & SCIM provisioning Exportable audit logs”
   • [claimed-docs] “Custom environments * Ramp schedules * Approval workflows * Advanced access control * SSO & SCIM provisioning * Exportable audit l…”

   proof ↗⚑ flag

Edge

1. platform engineerEvaluate flags at the edge (CDN workers or an edge/relay layer) close to users

   weight 2 · round drawn
   LaunchDarkly✓fullcommunity8/10

   LaunchDarkly documents a dedicated Cloudflare edge SDK (launchdarkly-docs-14) for evaluating flags directly in CDN workers, plus a self-hosted Relay Proxy (launchdarkly-docs-13/32, corroborated by community usage in launchdarkly-comm-2) that lets platform engineers run an edge/relay layer close to users for low-latency evaluation. Missing for 10: independent hands-on benchmarks of edge SDK latency/behavior and broader coverage of other edge platforms beyond Cloudflare.

   • [claimed-docs] “Cloudflare SDK reference”
   • [claimed-docs] “The LaunchDarkly Relay Proxy is a small Go application that runs on your own infrastructure.”
   • [claimed-docs] “The Relay Proxy”
   • [community] “LaunchDarkly clients download all flags in the project on connection, then stream changes. ... LaunchDarkly has an open source 'Relay' you c…”

   proof ↗⚑ flag

   GrowthBook✓fullclaimed8/10

   GrowthBook documents a dedicated Cloudflare Workers edge SDK that evaluates feature flags and Visual Editor experiments at the edge to cut flicker before HTML ships, plus a self-hostable Proxy layer that caches SDK payloads near users and keeps serving flags if the API is down. Missing for 10: independent hands-on evidence specifically of the edge/Cloudflare SDK in production (community evidence covers general SDK/self-host use but not edge-worker deployments) and broader edge-platform coverage beyond Cloudflare.

   • [claimed-docs] “Evaluate GrowthBook feature flags and Visual Editor experiments at the edge on Cloudflare Workers to cut flicker before HTML ships.”
   • [claimed-docs] “Run the GrowthBook Proxy next to your app to cache SDK payloads, reduce latency, and keep serving flags if the API is down.”
   • [claimed-docs] “The SDK Connection Endpoint provides readonly access to a subset of your feature flag data, just enough for the GrowthBook SDKs to assign va…”

   proof ↗⚑ flag

Evaluation

1. platform engineerMy server SDKs evaluate flags locally from a cached ruleset — microsecond decisions with no network call per flag check

   weight 3 · round to GrowthBook
   LaunchDarkly~partialcommunity6/10

   Community evidence confirms LaunchDarkly's core streaming architecture — SDKs download the full flag ruleset once and then stream updates, enabling local evaluation without a per-flag network call (launchdarkly-comm-2, launchdarkly-comm-9), and the Relay Proxy exists specifically to cache rulesets on-prem for resiliency (launchdarkly-docs-13, launchdarkly-docs-32). However there is no first-party SDK doc in this pack explicitly describing the local-evaluation/streaming-store mechanism or microsecond latency claims, and one community comment raises concern about network dependency per request (launchdarkly-comm-6), showing the story is not fully substantiated end-to-end. Missing for 10: dedicated SDK architecture docs describing the in-memory flag store/streaming model, explicit benchmark/latency numbers, and resolution of the community concern about request-time dependency.

   • [community] “LaunchDarkly clients download all flags in the project on connection, then stream changes. ... LaunchDarkly has an open source 'Relay' you c…”
   • [community] “Key benefits of LaunchDarkly: rules-based targeting so flags can be set at group level, custom rollout by % of users or attributes, centrali…”
   • [claimed-docs] “The LaunchDarkly Relay Proxy is a small Go application that runs on your own infrastructure.”
   • [claimed-docs] “The Relay Proxy”
   • [community] “It's a pretty question[able] architecture that outsources feature flags to a third party. You're creating a critical path hard dependency an…”

   proof ↗⚑ flag

   GrowthBook✓fullprobed9/10

   GrowthBook's core architecture is a cacheable JSON ruleset that SDKs evaluate locally, confirmed by docs (SDK Connection Endpoint providing readonly flag data for local assignment, in-memory attribute evaluation) and corroborated by a community founder comment describing exactly this model plus a runtime probe showing the keyless payload endpoint serving flag data for local SDK evaluation. Missing for 10: explicit published latency/microsecond benchmarks or independent third-party performance testing.

   • [claimed-docs] “The SDK Connection Endpoint provides readonly access to a subset of your feature flag data, just enough for the GrowthBook SDKs to assign va…”
   • [claimed-docs] “The actual values of targeting attributes (e.g., user IDs, emails) are never sent to GrowthBook. They are only stored in memory locally with…”
   • [community] “With GrowthBook, there's a single cacheable JSON definition of all of the features and then evaluation happens locally within the SDKs... Th…”
   • [probe] “PROBE runtime (recorded 2026-09-05): full KEYLESS self-host roundtrip on this machine — the official growthbook/growthbook docker image boot…”
   • [claimed-docs] “Run the GrowthBook Proxy next to your app to cache SDK payloads, reduce latency, and keep serving flags if the API is down.”

   proof ↗⚑ flag

Sdks

1. developerUse official SDKs across my whole stack — backend, web, and mobile — with consistent flag behavior

   weight 2 · round to LaunchDarkly
   LaunchDarkly✓fullcommunity8/10

   LaunchDarkly documents '30 idiomatic SDKs' across languages/platforms, OpenFeature provider support, edge/Cloudflare SDKs, and a Relay Proxy for consistent flag delivery, and community evidence independently confirms a working mobile client SDK and consistent streaming/flag-download behavior across SDKs. Missing for 10: an explicit enumerated list of backend/web/mobile SDK names in the pack and independent hands-on cross-platform consistency testing (community notes mobile SDK's MAU-based pricing as a practical caveat, not a functional gap).

   • [claimed-docs] “30 idiomatic SDKs”
   • [claimed-docs] “OpenFeature providers”
   • [claimed-docs] “The LaunchDarkly Relay Proxy is a small Go application that runs on your own infrastructure.”
   • [claimed-docs] “The Relay Proxy”
   • [claimed-docs] “Cloudflare SDK reference”
   • [community] “LaunchDarkly has a mobile client SDK, but they charge by MAU, which would be untenable. So folks tend to write a proxy for the mobile apps t…”
   • [community] “LaunchDarkly clients download all flags in the project on connection, then stream changes. ... LaunchDarkly has an open source 'Relay' you c…”

   proof ↗⚑ flag

   GrowthBook✓fullcommunity7/10

   GrowthBook offers 24 official SDKs spanning backend, web/React, and mobile (Android/iOS) per GitHub README, with docs showing consistent flag-evaluation patterns (useFeatureIsOn, useFeatureValue, targeting attributes evaluated locally) and community reports of quick multi-stack integration (Vue frontend in hours). An older 2021 community comment noted no native mobile SDKs yet, but this is superseded by the more recent GitHub claim explicitly listing Android/iOS SDKs, so it's not a live contradiction. Missing for 10: independent verification of identical targeting/rollout behavior across all 24 SDKs, and more recent hands-on confirmation of mobile SDK maturity/parity with backend/web SDKs.

   • [github] “24 SDKs including React, Python, Android, and iOS.”
   • [claimed-docs] “Get GrowthBook feature flags and experiments running in your app in a few lines of code, then pick a language-specific SDK.”
   • [claimed-docs] “const showNewFeature = useFeatureIsOn("new-feature");”
   • [claimed-docs] “const showNewFeature = useFeatureIsOn("new-feature"); const buttonColor = useFeatureValue("button-color", "blue");”
   • [claimed-docs] “Add targeting conditions to a rule to control which values a user sees.”
   • [community] “At Miggos we have implemented Growth book and we got FF and A/B in our Vue frontend in a few hours. Excited to see the progress of the platf…”
   • [community] “We don't have native mobile SDKs yet, but it's something we want to support in the future. Mobile is a little tricky since you either need t…”

   proof ↗⚑ flag

Standards

1. platform engineerUse the vendor through OpenFeature providers so my flag code isn't locked to one vendor's SDK API

   weight 1 · round to LaunchDarkly
   LaunchDarkly✓fullclaimed8/10

   LaunchDarkly documents official OpenFeature providers (launchdarkly-docs-33) alongside its 30 idiomatic SDKs, letting platform engineers write against the vendor-neutral OpenFeature API instead of a proprietary SDK surface. missing for 10: independent hands-on confirmation of OpenFeature provider behavior/parity across languages and any community discussion validating the OpenFeature integration in practice.

   • [claimed-docs] “OpenFeature providers”
   • [claimed-docs] “30 idiomatic SDKs”

   proof ↗⚑ flag

   GrowthBook—none0/10

   The evidence pack describes GrowthBook's own 24 SDKs, REST API, and proxy, but contains no mention of OpenFeature or an OpenFeature provider integration anywhere. Since this axis (avoiding vendor lock-in via OpenFeature) plausibly applies to a flagging platform, the absence of any supporting evidence yields 'none' rather than 'na'.

   • [github] “24 SDKs including React, Python, Android, and iOS.”
   • [claimed-docs] “Get GrowthBook feature flags and experiments running in your app in a few lines of code, then pick a language-specific SDK.”
   • [claimed-docs] “Get GrowthBook feature flags and experiments running in your app in a few lines of code”

   proof ↗⚑ flag

Streaming

1. developerFlag changes propagate to connected SDKs in seconds via streaming or fast polling — a kill switch actually kills

   weight 2 · round to LaunchDarkly
   LaunchDarkly✓fullcommunity7/10

   LaunchDarkly's SDK architecture is documented and independently corroborated as a streaming (SSE) model where clients download all flags then receive real-time updates, with an optional self-hosted Relay Proxy for resiliency — this is the mechanism that lets a flag toggle (kill switch) propagate quickly to connected SDKs. Community commentary explicitly calls out 'centralized feature flag service via SSE with minimal latency' as a key benefit, though some users note occasional API/SDK timeout issues that are latency-adjacent concerns rather than a contradiction of the streaming design. Missing for 10: first-party documentation with explicit propagation-time benchmarks (e.g. 'seconds'), explicit 'kill switch' terminology/case study, and hands-on confirmation of polling fallback speed.

   • [community] “LaunchDarkly clients download all flags in the project on connection, then stream changes. ... LaunchDarkly has an open source 'Relay' you c…”
   • [community] “Key benefits of LaunchDarkly: rules-based targeting so flags can be set at group level, custom rollout by % of users or attributes, centrali…”
   • [claimed-docs] “The LaunchDarkly Relay Proxy is a small Go application that runs on your own infrastructure.”
   • [claimed-docs] “The Relay Proxy”
   • [community] “My logs show frequent timeouts when looking up setting through their API. Their SDK default timeout is 3s. Wouldn't want to wait longer.”

   proof ↗⚑ flag

   GrowthBook~partialcommunity4/10

   GrowthBook's docs confirm a single cacheable feature-flag JSON payload that SDKs evaluate locally (growthbook-comm-6), a Proxy that caches payloads to reduce latency and keep serving flags if the API is down (growthbook-docs-11), and environment/rule toggles that could function as kill switches (growthbook-docs-3, growthbook-docs-22). However, no evidence in the pack documents actual streaming/SSE or fast-polling update latency in seconds, nor confirms a tested kill-switch effect at runtime. Missing for 10: explicit documentation of push/streaming update mechanism and propagation-time guarantees, and a hands-on/community test showing a kill switch immediately disabling a flag across connected SDKs.

   • [claimed-docs] “Run the GrowthBook Proxy next to your app to cache SDK payloads, reduce latency, and keep serving flags if the API is down.”
   • [community] “With GrowthBook, there's a single cacheable JSON definition of all of the features and then evaluation happens locally within the SDKs... Th…”
   • [claimed-docs] “Use environments such as dev, staging, and production to enable or disable feature flags independently.”
   • [claimed-docs] “How rules override a feature's default value — targeting users, rolling out gradually, experimenting — and how to scope, ramp, and schedule …”
   • [claimed-docs] “The SDK Connection Endpoint provides readonly access to a subset of your feature flag data, just enough for the GrowthBook SDKs to assign va…”

   proof ↗⚑ flag