What is an error tracking MCP server?
An error tracking MCP server lets an AI coding agent read production errors directly. It exposes an error tracker's data as tools over the Model Context Protocol, so Claude Code, Cursor or ChatGPT can list errors, read stack traces and often mark them resolved, with nothing pasted into the chat and no access to your servers.
How it works
MCP (Model Context Protocol) is an open-source standard for connecting AI applications to external systems. An error tracker already holds your production exceptions. The MCP server is the layer that lets an agent ask for them.
- Your app reports each exception to the tracker through an SDK.
- The tracker groups repeats into one error, with a count and first and last seen times.
- You connect the MCP server to your agent, with a URL and a sign-in or with a local command and a key.
- You ask "what broke today?" The agent calls a tool, reads the result, and opens the files in the stack trace in your repository.
What the tools do
Names differ by vendor. These are Runheed's, and most servers cover the same jobs:
| Job | Example tool | Question it answers |
|---|---|---|
| List | runheed_get_recent | What broke today? |
| Search | runheed_search_errors | Are there unresolved TypeErrors in checkout? |
| Read one error | runheed_get_error | What is the stack trace? |
| Read the surroundings | runheed_get_error_context | What was the request, release and trail of events? |
| Spot changes | runheed_get_trends | Did errors go up after yesterday's deploy? |
| Close the loop | runheed_resolve_error | This is fixed in commit abc123. |
Remote or local
| Compare | Remote (hosted) | Local (stdio) |
|---|---|---|
| Setup | A URL, then sign in | A command that runs on your machine, usually with a key |
| Works in claude.ai and ChatGPT | Yes | No |
| Works on a CI box | With a key in a header, where the server allows it | Yes |
| Sign-in | Usually OAuth, revocable | A token in a config file or environment variable |
Runheed has both: https://runheed.com/api/mcp with OAuth, and a local server for clients without sign-in.
What it is not
- It is not an error tracker on its own. It reads what the tracker recorded, so your app has to report errors first.
- It is not a log reader. It returns grouped exceptions, and log lines stay in your log tool.
- It does not fix anything by itself. The agent writes the fix, in your repository, under the permissions you gave it. See can an AI agent fix production bugs by itself.
- It is not access to your servers. A well-scoped server can read errors and change their status, and nothing else.
Which trackers have one
Runheed, Sentry, Honeybadger, Rollbar and BugSnag all have one, and they differ in where the server runs and what comes back. See error tracking MCP servers compared, which is not a full list. For one vendor in detail, see what Sentry's MCP server can do.
Runheed is built around this use: nine tools, one URL, and nobody has to notice an error, find it and paste it in. The tools are listed on the MCP server page.
Questions
- Do I need an MCP server to give an agent my errors?
- No. You can paste a stack trace, or give the agent shell access to logs. MCP lets the agent look errors up itself, in grouped form, without server credentials.
- Which agents can use one?
- Any MCP client. That includes Claude Code, Cursor, Codex, VS Code, claude.ai and ChatGPT. Local servers only work in desktop and CLI agents.
- Is it safe?
- It depends on the server's scopes and what your app sends. Runheed's tools read errors and change their status, secret headers are redacted when an event arrives, and any connection can be revoked in Settings. Details are on the is it safe page.
- How much does one cost?
- It depends on the tracker. Runheed's server is included in every plan, and plans start at $29/month.