Key takeaways
- Every MCP connection identifies its client automatically.
- Break every metric down by client.
- A tool can fail in one client and work in another.
Knowing which AI agents call your MCP server is one of the most useful things you can measure. Clients differ in how they discover tools, format arguments, and retry, so your client mix is a design input, not a vanity number.
How it's captured
Each MCP connection identifies its client. Analytics at the protocol layer records that automatically, so you can break every metric down by client without extra work.
What to look for
- Distribution: which clients drive most usage
- Growth: which client is growing fastest
- Per-client success: a tool may fail in one client and not another
Why it changes decisions
If completion is high in one client and low in another, the fix is client-specific. If growth concentrates in a single client, test changes against it first. Client-level data turns a vague 'it works' into something you can act on.
See this on your own server
TrackMCP turns your MCP server's calls into adoption, workflows, and outcomes. One line to install.