library · sentry
Sentry, where the goal is a count that should fall.
Sentry's MCP server finds issues, reads events and updates issue state. It is the clearest case for a measurable goal: unresolved issues in production is a number, a read tool returns it, and an agent that triages all night either moved it or did not.
Example, not real data.
What the agent can reach.
From Sentry's documentation as of 2026-09-02. The catalog shows what it publishes when it connects.
What changes when it runs through the gateway.
A triage preset allows find_issues, get_issue_details and update_issue and nothing else, so the agent can resolve, ignore and assign, and cannot create projects or teams. Each update is logged with the reason the agent gave, under the name of the person it ran for. If resolving a certain kind of issue keeps being followed by the count going up, the next such update is refused with that evidence.
One caveat.
The goal above works only if find_issues counts as read-only, and that is decided by Sentry's own annotations; a library reference cannot add names. If Sentry marks nothing, add the same URL as a custom server and list the read tools there. Everything else on this page holds either way.