MCP
Your agent reads the bug report before you have read it.
You already write code with Claude Code or Cursor. The bug report from a real user is sitting in a different tab, and the ritual is to copy the description, then the stack trace, then the browser, then explain the steps. The MCP server removes the ritual: the agent asks for the report and gets all of it.
Start freeNo credit card required.
in your editor
> list_bugs since=7d
a3f91c04 bug pending 2h Checkout fails on the discount step
TypeError: Cannot read properties of undefined
> get_bug a3f91c04
Chrome 131 · macOS 15.1 · 1440 × 900 · Europe/Madrid
TypeError: Cannot read properties of undefined (reading 'total')
at CartSummary (app.4f2c1.js:284:19)
Steps 1 /cart → 2 click "Apply code" → 3 submit form#discount
Screenshot https://userwants.app/s/a3f91c04Three tools
What your agent can actually do
Deliberately small. A broad surface is how an agent ends up doing something surprising in a system that talks to your customers.
- list_bugs
- Reports newest first, filtered by type, status, age or free text over the title and description. Each row is one line, ending with the first JavaScript error if one was captured.
- get_bug
- One whole report from an 8-character ref: description, reporter, page, browser and viewport, the JavaScript errors with stack traces, the numbered steps, and a screenshot link.
- set_bug_status
- Moves a report through pending, planned, in progress, completed and rejected. Only registered for owners and admins, and it does not email the reporter unless the call asks it to.
There is also a triage_bugs prompt, which appears as a slash command in editors that support them.
Access
OAuth, and the project is in the URL
Each project has its own endpoint, so the project lives in the URL rather than in the credential and the tools take no project argument — an agent cannot address a project it was not given. The first connection opens a browser for your approval; after that the agent acts as you, with exactly your role permissions. The full reference is in the MCP docs, and the per-editor configuration is on the setup page.
Why it matters
The context is the whole job
An agent given “checkout is broken” writes a plausible guess. An agent given the exception, the file and line, the browser, the viewport and the three clicks that led there writes the fix. Everything in that second list is collected automatically, which is what makes it worth exposing over a protocol at all — more on stack traces from real users.
FAQ
Questions
- What is an MCP server for bug reports?
- MCP is the protocol AI coding tools use to reach outside data. An MCP server for bug reports lets your agent pull a real user report — description, stack trace, steps, browser — into its context directly, instead of you pasting it in.
- Which editors can connect to it?
- Anything that speaks HTTP MCP: Claude Code, Cursor, VS Code and Zed among them. The configuration is a URL and nothing else, so what differs between editors is only which file that URL goes in.
- Is there a token to manage?
- No. Access runs through OAuth, so your editor opens a browser once and you approve it. There is no key to create, rotate or accidentally commit, which is why the config snippet is safe to check into your repository.
- Can the agent change things in my account?
- Only what you could change yourself. It acts as you, so it sees the projects you are a member of, and the status tool is not even registered unless your role is owner or admin.
- Does it email my users by accident?
- No. Notifying the reporter when a status changes is off unless the call explicitly asks for it. Sending mail to somebody else customer is not a side effect any agent should be able to trigger unprompted.
Point your editor at your own inbox.
One URL in your MCP config, one browser approval, and your agent can read what your users reported.