The metrics that matter for an MCP server

Key takeaways

Lead with adoption and outcomes, not raw call volume.
Reliability metrics support the story; they don't define it.
A fast server nobody completes a task on is not healthy.

It is easy to drown an MCP dashboard in engineering metrics. Latency, throughput, error codes. They matter, but they are not the story. Here are the numbers that tell you whether the server is doing its job.

Adoption

Active clients: how many distinct clients connect
New vs returning connections: is usage growing and sticking
Client mix: Claude, Cursor, ChatGPT, custom agents

Behavior

Tool adoption: which tools get called, which are ignored
Most common workflow: the path agents actually take
Where sessions stop: the step before people give up

Outcomes

A few tools do almost all the work; the long tail barely moves.

Completion rate: sessions that reach a useful result
Returning usage: clients that come back within a week

Reliability (supporting)

Silent failures: errors inside a 200 OK
p95 latency per tool: the tail that ruins sessions

Lead with adoption and outcomes. Keep reliability close, but do not let it define the product. A fast server nobody completes a task on is not a healthy server.

See this on your own server

TrackMCP turns your MCP server's calls into adoption, workflows, and outcomes. One line to install.

Get started