What it is for
Errors arriving inside your editor. Instead of opening the dashboard to investigate, you ask the agent already open there: "what is breaking in production?", "why does this error happen?".
It is an MCP (Model Context Protocol) server built into meulog. Claude Code, Cursor, Zed and any MCP client talk to it.
The four tools
list_errors: the project open errors, most frequent first. Takes a limit, the last N hours, and a status.show_error: everything about one error. Stack, context, volume, affected versions, and which deploy it debuted in.analyze_error: asks the AI analyst configured in meulog for the likely cause, with the code around the line that broke (when a repository is on file).mark_error: changes the status. Fixed requires the version that fixed it, same as on screen.
How to wire it
A POST to /mcp, with the project server token in the
Authorization header. It is JSON-RPC 2.0 over HTTP, with no SSE and no
session: every tool is question and answer.
A project token reaches that project only, and mark_error is the only
tool that writes. The agent does not delete, does not merge and does not touch
settings.
Why this exists here
Sentry and Rollbar have it too. The difference is analyze_error: meulog
analysis runs on the analyst you configured. It can be Claude, it can be
OpenAI, and it can be an Ollama on your own machine, if the error cannot leave the
building. It costs nothing per third-party token.