Hand a production error to your AI agent over MCP

Fixing a production error with an AI agent usually starts with copying. The stack trace from your error tracker, the request, the steps that led to it, a note about which release is live — pasted into a chat, piece by piece, until the agent has enough to guess.

Most of that context already exists. It just lives in the wrong place.

Give the agent the error, not a summary

AxoWatch is an MCP server. Once your agent is connected, it can read your sites’ uptime, page speed, flows and app errors directly — and, when you ask, take an error as a ready task.

Connecting takes one command and one sign-in. In Claude Code:

claude mcp add --transport http --scope user axowatch https://<your-axowatch>/mcp

The first time the agent uses it, a browser window opens and you sign in to AxoWatch. There are no API keys to copy, store or rotate. Codex, Cursor and other agents that speak MCP connect the same way, each with its own command on the connect screen.

Press Fix with AI

On any error in AxoWatch there is a Fix with AI button. It gives your agent a short instruction: call the errors_fix_task tool for this error and do what the task says.

The task is where the context lives. It carries:

• The error — how often it happened, since when, in which environment and release.
• The request that hit it.
• The stack with your source lines, at the revision that is deployed.
• What the app did before it broke — the queries and steps leading up to it, newest first.
• How to tell what kind of failure it is. Not every error is a code bug: some are missing data an admin can fill in, some need a migration, some are the environment. The task asks the agent to find out first, and fix the cause where it lives.

Everything a visitor typed — messages, request bodies, page text — arrives marked as untrusted data. The agent reads it as evidence, never as instructions.

Your agent works, you approve

The agent works in your repository, with your tools, the way it always does. It reproduces the error, changes the code, runs the tests and proposes the change. You review it and merge it — AxoWatch never touches your code.

The same works for slow pages. A Fix with AI on a page that got slower hands the agent the failing Lighthouse audits, worst first, and the exact profile we measure with, so it can prove the fix with the same measurement — not by gaming the score.

Closed with a note

When the fix is live, the agent marks the error resolved. It can’t do that silently: resolving requires a note on what changed, and the note is signed with the agent’s name.

Claude Code · Resolved

An empty cart had no total, so checkout failed. The cart now returns a zero total; covered by a test.

If the same error happens again, it is reopened and flagged as back — with the note right there, so whoever looks next knows what was tried.

What the agent can and can’t do

Connected over MCP, an agent can read the uptime, speed, flows and app errors of your sites. The only thing it can change is an error’s state — resolved or ignored — and only with a note. It acts as you, and you can disconnect it at any time; the notes it left stay, signed with its name.

Where to start

Connect the agent you already use, open your noisiest error and press Fix with AI. Read the task it gets before it starts — it is the clearest bug report you have ever written, and you didn’t have to write it.