- Completion rate is the one metric that can't be gamed by busy failure.
- Tie 'complete' to a real outcome where you can.
- Move it by fixing the tool that ends the most sessions early.
If you could keep only one metric for your MCP server, make it workflow completion rate: of the sessions that start, how many reach a useful result. Almost everything else is a lever on this number.
Why it beats call volume
Call volume rewards busy failure. A server where agents call tools constantly and never finish looks active and is broken. Completion rate cannot be gamed that way.
How to define 'complete'
Where you can, tie completion to a real outcome: an issue created, a query answered, a message sent successfully. Where you cannot measure business value directly, use an honest proxy like a session that ends on a successful terminal tool rather than a retry loop.
Moving it
- Fix the tool that ends the most sessions early
- Make schemas forgiving so valid intents don't fail
- Improve descriptions so agents pick the right tool sooner
Watch completion rate weekly, and let it decide what you fix first.
See this on your own server
TrackMCP turns your MCP server's calls into adoption, workflows, and outcomes. One line to install.