Query MCP Analytics over MCP

PostHog's MCP server exposes MCP Analytics as a set of tools, so you can investigate how agents use your server from Claude Code, Cursor, or any other MCP client – without switching to the web app. It's the fastest way to ask "which of my tools is failing, and why?" while you're already in your editor.

Setup

Follow the MCP server setup to connect PostHog to your client. MCP Analytics tools need the mcp_analytics:read scope, and the ones that write data (feedback, missing capabilities, recomputing clusters) need mcp_analytics:write.

What you can do here

Explore sessions

• List MCP sessions and filter them down
• Open a single session and read its tool calls in order
• Generate an intent summary for a session

Understand intent

• Retrieve intent clusters and their tool distribution
• Recompute clusters after new data lands

Dig into tool quality

• Per-tool stats: volume, error rate, and latency
• Daily trend for a tool over time
• Recent failures, with error type and message
• Top users of a given tool
• Neighboring tools that get called alongside it
• Sample intents behind its calls
• Harness and client breakdown

Send feedback

• Submit feedback about a tool
• Report a missing capability your agent needed but couldn't find

Example prompts

Point your agent at these once the tools are connected:

• "Which of my MCP tools has the highest error rate this week?"
• "Show me the last 10 failed calls to search_docs and what the errors were."
• "What are agents actually trying to do when they call create_issue? Summarize the intent clusters."
• "Which tools get called right before run_query?"

Related

• Explore the same data visually in the web app.
• See the full list of PostHog MCP tools in the MCP docs.
• Act on Self-driving reports about failing tools in PostHog Desktop.