Capture the bug. Let AI fix it.

AI bug fix: give your agent
everything it needs to fix it

AI can fix a bug it can actually see. TraceBug captures the browser session — replay, console errors, the failed request, the exact repro steps — into one file your coding agent reads over a local MCP server. It then reproduces and fixes the bug, with a generated failing test to prove it.

Watch AI fix a real bug Add to Chrome — free

How AI bug fixing works with TraceBug

1
Capture
One shortcut records the session — replay, console errors, network, a repro timeline — into one offline .html file.
2
Hand to your agent
A local MCP server exposes the report to Claude Code, Cursor, or Windsurf. No upload, no dashboard.
3
AI fixes it
The agent reads the evidence, finds the root cause, and patches the code — using the failing test to know when it's done.

Why pasting a screenshot into ChatGPT doesn't fix the bug

An AI agent can only reason about what it's given. A screenshot and “it's broken” forces it to guess. The evidence that actually explains a bug — the console stack trace, the 500 response body, the click that triggered it, the state at the moment it failed — exists at capture time and usually never reaches the model. TraceBug captures all of it and hands it over structured, so the agent debugs from facts instead of guessing.

The best MCP setup for fixing bugs

TraceBug ships a local, stdio-only MCP server — one command, zero network connections:

$ claude mcp add tracebug -- npx -y tracebug mcp --dir ./bug-reports

It gives your agent read-tools over the report (console errors, network activity, repro steps, screenshots) plus a generated failing Playwright test the agent runs to reproduce the bug and verify its fix. Setup guides for Claude Code, Cursor, and Windsurf.

From root cause to a verified fix

Diagnosis isn't a fix. Every TraceBug report embeds a runnable failing test that replays the captured session and asserts the failure is gone — red while the bug exists, green once it's fixed. The agent runs it, patches, and re-runs until green: the report stops being evidence to read and becomes something to iterate against.

Run the generated test → red confirms the bug → fix → green proves it's gone.

Local, not cloud. Capture runs in your browser, the MCP server runs on your machine over stdio with zero network connections, and secrets are masked at capture. Your bug data — and your code — never leave your machine. MIT open source.

See it fix a real bug
Unedited transcript of Claude going from crash to root cause, then a verified fix.
Watch the proof