posthog-ff-mcp

A tiny, read-only MCP server that lets Claude read PostHog feature-flag state and metadata. It wraps PostHog's REST API — listing flags, reading a flag's full definition, checking staleness, and explaining how flags evaluate for a person. It never creates, updates, or deletes anything.

It's a single Python file (server.py) with inline dependencies, so it runs locally with uv and no build step.

Prerequisites

• uv installed.

• A PostHog personal API key scoped to at least feature_flag:read for your project. Create one at <your-host>/settings/user-api-keys.

Related MCP server: cc-plugin-mcp

Configuration

All config is via environment variables (no defaults):

See .env.example.

Run it locally

POSTHOG_HOST=https://eu.posthog.com \
POSTHOG_PROJECT_ID=13105 \
POSTHOG_PERSONAL_API_KEY=phx_your_key \
uv run server.py

The first run resolves dependencies automatically; the server then speaks MCP over stdio.

Add to Claude Code

claude mcp add posthog-ff \
  --env POSTHOG_HOST=https://eu.posthog.com \
  --env POSTHOG_PROJECT_ID=13105 \
  --env POSTHOG_PERSONAL_API_KEY=phx_your_key \
  -- uv run --directory /ABSOLUTE/PATH/TO/posthog-ff-mcp server.py

Replace /ABSOLUTE/PATH/TO/posthog-ff-mcp with this repo's absolute path.

Add to Claude Desktop

Add this to claude_desktop_config.json (macOS: ~/Library/Application Support/Claude/claude_desktop_config.json):

{
  "mcpServers": {
    "posthog-ff": {
      "command": "uv",
      "args": ["run", "--directory", "/ABSOLUTE/PATH/TO/posthog-ff-mcp", "server.py"],
      "env": {
        "POSTHOG_HOST": "https://eu.posthog.com",
        "POSTHOG_PROJECT_ID": "13105",
        "POSTHOG_PERSONAL_API_KEY": "phx_your_key"
      }
    }
  }
}

Restart Claude Desktop afterward.

Tools

Example prompts

• "List the active feature flags in PostHog."

• "Which flags are stale?"

• "Show me the full targeting for the new-onboarding flag."

• "Is beta-dashboard safe to delete?"

• "How do the flags evaluate for distinct id user_123?"

Scope & limitations

• Read-only. No create/update/delete — by design.

• Rate limits are team-wide (480/min, 4800/hr); the server surfaces 429s clearly.

• Not built in this version (possible future extensions): remote-config payloads, activity/version history, SDK local_evaluation, and the runtime /flags evaluator (which uses a different host and the project token instead of the personal key).