Sentry MCP server: setup, tools and scoping for Claude Code and Cursor

Most debugging sessions start in Sentry: an issue spikes, someone opens the stack trace, finds the release, looks at breadcrumbs, and then goes to the code. The Sentry MCP server lets an AI agent do that loop from inside Claude Code or Cursor, with the stack trace and the repository side by side. This guide covers setup, scoping, the tools, how to limit them, and what to watch with error data. If MCP is new to you, start with what an MCP server is.

Facts are from Sentry's MCP documentation and the getsentry/sentry-mcp repository as of October 2026.

What the server is

Sentry's MCP server is official and maintained by Sentry. It is still pre-1.0 (the latest release in early October 2026 was 0.47), so expect tool names to keep moving. Two ways to run it:

For sentry.io, use the hosted server. Use the local one for self-hosted Sentry, or where your policy requires the server process to run on your own machines.

Scope it to a project

The hosted URL can be narrowed:

• https://mcp.sentry.dev/mcp: all organizations you can access
• https://mcp.sentry.dev/mcp/<org>: one organization
• https://mcp.sentry.dev/mcp/<org>/<project>: one project

Sentry recommends project scoping. Tools default to that project, the discovery tools are hidden, and the agent stops asking which of your forty projects you meant. It is also the simplest least-privilege step there is.

Setup in Claude Code

claude mcp add --transport http sentry https://mcp.sentry.dev/mcp/acme/checkout-api

Run /mcp in a session to complete the OAuth login. Sentry also publishes a Claude Code plugin, which adds a subagent that handles Sentry work so its output does not fill your main context:

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

General steps for remote servers are in our Claude Code MCP guide.

Setup in Cursor and VS Code

Cursor: add the URL in Cursor's MCP settings and sign in when prompted. See our Cursor MCP guide.

VS Code: Command Palette, MCP: Add Server, choose HTTP and paste the endpoint.

Without OAuth (CI or a headless machine), the hosted server accepts a token in the header Authorization: Sentry-Bearer <token>.

Self-hosted Sentry

Run the local server and point it at your instance:

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

The user auth token needs the scopes org:read, project:read, project:write, team:read, team:write and event:write. Add --insecure-http only for an instance without TLS. Seer tools are off by default here, because self-hosted Sentry has no Seer.

The natural-language search_* tools translate a question into a Sentry query using an LLM. Run locally, they need your own provider through EMBEDDED_AGENT_PROVIDER (OpenAI, Azure OpenAI, Anthropic or OpenRouter) plus its API key. If you run a model in an EU Azure region, that is where to point it.

The tools

Read

Write

AI: analyze_issue_with_seer asks Sentry's Seer to analyse an issue and propose a root cause.

Limit it with skills

Sentry groups tools into skills: inspect, triage, project-management, seer and docs. There is no single read-only switch; you pick skills instead.

# Hosted: investigation only
https://mcp.sentry.dev/mcp/acme/checkout-api?skills=inspect,docs

# Hosted: everything except Seer
https://mcp.sentry.dev/mcp/acme/checkout-api?disable-skills=seer

Locally, use --skills and --disable-skills, or the MCP_SKILLS and MCP_DISABLE_SKILLS environment variables. Leaving out triage and project-management keeps the agent from changing issues, projects, DSNs or alert rules, which is the right default for most developers. Give triage to the people on the rotation.

What people use it for

• From issue to fix. In the repository: "Look at Sentry issue CHECKOUT-4F2, read the stack trace and breadcrumbs, find the cause in this code and propose a fix." This is the workflow that makes the server worth having.
• Release check. "Which new issues appeared since release 2026.10.1, and how many users do they affect?"
• Triage. With the triage skill: "Assign the top five unassigned issues in checkout-api to the owners in CODEOWNERS and add a note with the suspected cause."
• Alert hygiene. "List alert rules in this project with no recent triggers."

A standard debugging procedure belongs in a skill so every engineer's agent works through issues the same way.

Error data is sensitive data

Sentry events often carry more than stack traces: user IDs, emails, request bodies, headers, breadcrumbs with form input. When the agent reads an event, that content becomes part of the conversation and goes to your model provider. Two things help:

• Scrub at the source. Sentry's own data scrubbing and beforeSend filtering keep personal data out of events in the first place, which protects every consumer, not only the agent.
• Treat event content as untrusted. Error messages and breadcrumbs can contain text an attacker controlled. Do not let the same session send data out of the company without approval. See MCP security best practices.

The hosted server routes each organization's calls to its own Sentry region based on the organization's settings. Sentry's docs do not describe a separate EU endpoint for the MCP server itself.

Running it for a team

For a whole engineering organization you want the same skills enabled, the same write policy and one log across Sentry, GitHub, Datadog and the rest. That is the job of an MCP gateway. Walma AI Hub runs approved MCP servers like Sentry behind one policy and one audit log, with the models in the customer's own Azure tenant in an EU region. Related guides: GitHub MCP, Datadog MCP and Linear MCP.