Sentry MCP: production errors inside your coding agent

When something breaks in production, the routine is to open Sentry, read the stack trace and breadcrumbs, work out which release introduced it, then go back to the editor. Sentry MCP wires that path into the coding agent: it queries errors, issues, traces and logs, then uses the stack trace to find the code. Sentry says plainly that the service targets human-in-the-loop coding agents, so the tools follow the debugging workflow rather than covering everything Sentry does. The repository has been renamed from getsentry/sentry-mcp to getsentry/toolkit.

What it does

• Natural-language search: search_errors, search_issues, search_traces and search_logs translate questions into Sentry’s query syntax. These need an LLM provider configured.
• Skill groups: tools are grouped into skills such as inspect, triage, seer and project-management, and --skills or --disable-skills narrows the set.
• Seer root-cause analysis: available on sentry.io. Self-hosted Sentry has no Seer, so that group is dropped by default there.
• Claude Code plugin: adds a sentry-mcp subagent that Claude delegates to whenever you ask about Sentry errors.

Who it’s for

• Teams on Sentry who want Claude Code or Cursor to go straight from an alert to a fix.
• Self-hosted Sentry users who want agents to reach their internal error data too.

Setup

For sentry.io, any client with remote MCP support connects to https://mcp.sentry.dev/mcp and authorizes with OAuth. Claude Code users can install the plugin:

claude plugin marketplace add getsentry/sentry-mcp
claude plugin install sentry-mcp@sentry-mcp

For self-hosted Sentry, use the stdio transport with a User Auth Token that has the required scopes, plus --host:

npx @sentry/mcp-server@latest --access-token=TOKEN --host=sentry.example.com

Our take

Sentry is one of the most widely used error-tracking platforms, and its official server needs nothing deployed for sentry.io users. The tool selection is restrained and centered on reading errors and finding causes, rather than dumping dozens of admin endpoints on the agent. Two caveats. The token scopes listed for stdio include project:write, team:write and event:write, so it isn’t read-only; --disable-skills=project-management turns off the management tools if you don’t need them. And the natural-language search tools need a separate OpenAI, Anthropic or similar key; without one they’re unavailable while the rest keep working. Licensed FSL-1.1-Apache-2.0, a source-available license that converts to Apache-2.0 after two years.