| name | sentry |
| description | How to use the Sentry connector's tools — finding orgs/projects, searching issues and events, reading issue details, stack traces, traces and spans, running Seer root-cause analysis, and reaching long-tail operations (releases, alerts, monitors, dashboards, docs) via the tool catalog. Use when the user works with Sentry, errors, exceptions, crashes, stack traces, performance, spans/traces, releases, alerts, monitors, or asks to debug a production issue or investigate an error. |
You are using the Sentry connector (remote MCP). Its tools act on the
connected Sentry account's organizations and projects, and are read-only —
they observe and analyze, they do not mutate Sentry state. Apply this guidance
whenever you touch Sentry.
Resolve org and project first
- Almost every tool takes an
organizationSlug (and often a project). These
are slugs, not display names — resolve them before doing real work.
- Start with
find_organizations to list orgs the user can access, then
find_projects for a project slug. Don't guess a slug.
- If the user writes
my-org/my-project, that's organizationSlug/projectSlug —
parse it directly, no lookup needed. Project params accept a slug or numeric id.
whoami confirms who the token is acting as when access is ambiguous.
Two tiers of tools: direct and catalog
- A handful of operations are direct tools:
find_organizations,
find_projects, search_issues, search_events, analyze_issue_with_seer,
and get_sentry_resource.
- Everything else is reached through the catalog: call
search_sentry_tools with keywords to discover the right tool and its
schema, then execute_sentry_tool with that name and arguments. This is
how you get to issue details, stack traces, releases, alerts, monitors,
dashboards, profiles, tag distributions, issue activity, and docs.
- Don't assume a tool exists by guessing its name. If you don't see a direct
tool for the task, search the catalog first before deciding it's
unavailable — many useful operations are intentionally not top-level.
Searching: issues vs. events
search_issues returns grouped problems ("the error", deduped). Use it to
find what's broken — is:unresolved is:unassigned, level:error firstSeen:-24h.
search_events returns individual occurrences/spans/logs and supports
aggregations (counts, sums, averages). Use it for how many / when / where
and to drill in by time, environment, release, user, or trace id. Pick the
dataset (errors, logs, spans, metrics, profiles, replays) for the question.
query accepts natural language, Sentry search syntax, or a mix — an
embedded agent translates it but preserves explicit Sentry tokens you pass
(is:, level:, release:, environment:, firstSeen:-24h). State the
filter you want; you don't have to hand-write the whole syntax.
- These AI search tools depend on an LLM provider configured server-side — if
search returns a config error, fall back to explicit Sentry syntax.
Investigate one issue
- For a specific issue or a Sentry URL,
get_sentry_resource is the fastest
path: pass the whole URL and it auto-detects the type (issue, event, trace,
span, replay, breadcrumbs, monitor, snapshot). For traces it returns a
condensed overview by default.
- Via the catalog:
get_issue_details (by PROJECT-123 id or URL) for title,
culprit, counts, and latest event; get_event_stacktrace for a full thread
stacktrace; get_issue_tag_values for tag distributions (which URLs/browsers/
releases are affected); get_issue_activity for prior comments and who acted.
analyze_issue_with_seer runs Sentry's AI root-cause analysis and returns
a fix recommendation. Only run it when the user asks for analysis or you can't
determine the cause from the issue details — not as an automatic follow-up. A
fresh run takes ~2–5 minutes; set expectations and don't poll tightly.
Existing analyses return instantly, so reuse one when available.
Care
- This connector is read-only — it cannot resolve, ignore, assign, comment,
or create projects/teams/DSNs. If the user asks for those, say so and point
them to the Sentry UI rather than inventing a tool or claiming it was done.
- A "not found" issue/project is often a permissions or wrong-slug problem,
not a missing object — re-resolve the slug or say it isn't accessible rather
than inventing one.
- Large result sets: narrow with the
query parameter (orgs/issues/events all
cap results) and the time range instead of fetching broadly, then summarize.
- Quote stack traces and Seer output as evidence; don't paste an entire event
payload back verbatim. Stack frames and event data can contain user PII —
surface only what's needed to explain the issue.