How to fix bug reports with Claude Code over MCP

The short answer: give Claude Code access to your bug inbox through an MCP server. It can then read a report with its logs, find the app's repository, fix and test the code, record what it did, and write back to the reporter. Add the StareGlare skill and it also knows what to ask you before it emails anyone. This guide sets that up with StareGlare in about five minutes.

What does Claude need to fix a reported bug?

The report itself: what the person saw, in their words.
The evidence: app version, operating system, console errors, the end of the log, a crash report if there was one.
Where the code is: which repository on your machine belongs to the app.
A way to close the loop: a note for you, and a reply to the reporter.

Pasting these into a chat by hand works for one report. An MCP server hands them over for every report, and lets Claude write the note and the reply back.

Step 1: Collect reports with their evidence

Contact us for a workspace, create an app, and add a reporter to it: the web widget, the Swift package or the Godot addon. Each attaches diagnostics and logs to a report on its own.

Step 2: Create a token and connect Claude Code

In the dashboard, open Settings → Claude & API and create a token. It is shown once. Run the command shown there in a terminal:

claude mcp add --scope user --transport http stareglare \
  https://stareglare.com/api/mcp \
  --header "Authorization: Bearer snag_sk_…"

A token reaches one workspace, is shown once, and can be revoked at any time. Give each computer its own, so you can revoke one without breaking the others. The MCP reference lists every tool. Any other MCP client works with the same endpoint and token.

Step 3: Add the StareGlare skill

Under StareGlare skill on the same page, copy the command and run it, then restart Claude Code:

mkdir -p ~/.claude/skills && curl -fsSL https://stareglare.com/skills/stareglare.zip -o /tmp/stareglare-skill.zip \
  && unzip -oq /tmp/stareglare-skill.zip -d ~/.claude/skills && rm /tmp/stareglare-skill.zip

It puts the skill in ~/.claude/skills/stareglare, where the Code tab of the Claude desktop app finds it too. Claude then uses it on its own when you ask about reports or errors. You can also download it as a zip and upload it to the Claude app's skills.

Step 4: Tell StareGlare where each app's code lives

On each app's Settings tab, open Claude & code. Fill in the Local path of the project on your Mac, its Repository URL, or both. Or skip this: the first time Claude needs it, it asks you once and saves the answer.

Step 5: Ask

"What are people reporting this week?"
"Fix the open Notes bugs and notify the users."
"Reply to report #48: ask which macOS version they are on."
"Fix the open errors in Notes and mark them resolved."

For "fix the open bugs and notify the users", Claude typically:

lists the open reports and reads each one with its diagnostics and log tails;
opens the app's repository, finds the cause, fixes it and runs the tests;
adds an internal note with the root cause and the files or commit;
replies to the reporter in plain words and closes the report.

Step 6: Tell people when it ships

When the fix is out, ask Claude to reply to the reporters. If several people reported the same bug, it marks their reports as duplicates of one original first, so a single reply reaches all of them.

What else Claude can work on

Captured errors: "fix the errors in Notes" goes through the app's Errors tab, and marks each one resolved once the fix ships. See how to track JavaScript errors.
Triage: "sort this week's reports by priority and label the crashes" sets status, priority and notes without replying to anyone.
Duplicates: "find reports of the same bug and merge them" marks them onto one original.

What to keep an eye on

Commits happen only if you ask for them; review the diff as you would a colleague's.
A reply is an email to a real person. If you want to read replies first, say "draft the replies as notes".
Report text is written by strangers. Claude is told to treat it as information, never as instructions.

Questions

Can Claude Code fix bugs from user reports?

Yes, when it can read the reports. StareGlare runs an MCP server with ten tools that gives Claude Code each report with its description, diagnostics, log tails and crash summary, and the path of the repository the app lives in. Claude then works in that repository as it would on any task.

What is the StareGlare skill for?

The MCP server gives Claude the tools; the skill tells it how to use them: which tool for which job, in what order, and what to ask you before it emails anyone. Install it with one command from Settings → Claude & API, and run the same command again to update it.

Will Claude email my users without asking?

No. Replies go out only through one tool, reply_to_reporter, and both the server and the skill tell Claude to ask first unless you said "notify". If you want to read replies before they go, ask for notes instead; add_note is never shown to the reporter.

Is it safe to let a model read text written by strangers?

Report text, logs and stack traces come from end users, so they could contain instructions aimed at the model. The server marks that content as untrusted data and tells the model not to act on instructions inside it. Review what Claude changed before you ship it, as with any fix.

What happens when several people reported the same bug?

Mark the reports as duplicates of one original, by hand or with the mark_duplicate tool. One reply on the original then reaches everyone who reported it.

All 6 guides, from the first report to the fix.