Sentry error triage by an AI employee on Claude Code or Codex

What the employee does

• Reads unresolved Sentry issues that are new or regressed since its last run.
• Deduplicates against your board, so a noisy error gets a comment, not ten new issues.
• Files a useful issue: impact, first and last seen, the release it started in, the top stack frames and a likely cause, clearly labelled as a guess.
• Sets a priority from rules you write, such as anything that blocks sign-up or payment is urgent.
• Fixes the obvious ones: a null check, a missing translation key, a bad redirect. It opens a pull request with a test and waits for review.

How to set up Sentry triage on Claude Code or Codex

1. Connect Sentry's MCP server

   Sentry hosts an MCP server at mcp.sentry.dev with OAuth, so there's nothing to install and no token to paste. Add it, start the agent and sign in to Sentry when it asks. It gives the agent Sentry's issues, errors, projects and Seer analysis.

   # Claude Code: add Sentry's hosted MCP server, then sign in with OAuth
   claude mcp add --transport http sentry https://mcp.sentry.dev/mcp
   
   # Codex: through the mcp-remote bridge
   codex mcp add sentry -- npx -y mcp-remote@latest https://mcp.sentry.dev/mcp

2. Copy the role brief

   Into CLAUDE.md for Claude Code or AGENTS.md for Codex, in the repository the errors come from, so it can read the code behind each stack frame.

   # Role: error triage
   Runs from the recurring issue "Sentry triage", weekdays at 9:00 and 14:00.
   
   1. In Sentry, list unresolved issues first seen or regressed since the
      last run, in the production environment.
   2. For each: check the board for an existing issue (search the Sentry
      issue ID). If there is one, add the new count as a comment.
   3. Otherwise open an issue: title from the error, users and events
      affected, first and last seen, the top stack frames, the release
      it started in, and your best guess at the cause, marked as a guess.
   4. Priority: Urgent if it blocks sign-up, payment or data; High if it
      affects more than 1% of sessions; otherwise Medium.
   5. If the fix is small and obvious, take it as your next issue and
      open a pull request with a test. Don't merge.
   6. Never resolve, ignore or delete anything in Sentry yourself.

3. Put it on a schedule

   Make triage a recurring issue, for example --repeat "weekdays at 9:00", and a second one in the afternoon if your error volume needs it.

4. Review the fixes

   Fix pull requests go through your normal review. When the employee is unsure about impact or ownership, it marks the issue Needs action and asks.

Source: Sentry's MCP server docs

Rules that keep it safe

• Read Sentry, write to the board. Resolving, ignoring or deleting in Sentry stays with people.
• Production only: point it at your production environment so it doesn't file staging noise.
• A guess is a guess: the cause in each issue is labelled as a hypothesis until someone confirms it.
• No merges: fixes are pull requests with a test, never direct pushes.
• Personal data: Sentry events can include user details; keep the issue to counts and IDs, not emails or IPs.

Step by step for Claude Code and Codex

The built-in commands, the tools to connect and the schedule for each agent: Sentry with Claude Code and Sentry with Codex.

An engineer that starts the day with your errors

Triage is a natural recurring issue for an AI software engineer. Crewmark turns your Claude Code or Codex into that employee, with working hours and a heartbeat.

1. Start from the role

   The AI software engineer takes issues from a board and opens pull requests.

2. Add triage as a recurring issue

   It runs triage when it's due, then works through the issues it filed.

3. More engineering jobs

   See Claude Code for engineering and AI code review on every pull request.

FAQ