Let the agent read the code

This tool gets you as far as what changed. Working out what that code does is a different job, so an MCP server ships with it and the agent reads the results itself.

One line and you're done

claude mcp add gitmoru -- npx tsx <path>/src/mcp/index.ts

Hit "Hook up an agent" in the app and it prints the real config path on your machine.

Works with

Claude Code
Claude Desktop
Codex CLI
Gemini CLI
Cursor
Windsurf
VS Code

Eleven tools

scanStart here. Returns a caseId
open_caseOpen a case the app already scanned
triageWhat to read first
list_changesEvery changed file
diff_fileThe file before the attack against the file now
read_fileThe whole file
list_forced_pushesWhich branch lost how many commits
check_accessDeploy keys, webhooks, pending invites
preview_restoreWhat would be restored where
share_summaryA summary to paste in a team channel
list_casesEverything on disk

What comes back

list_forced_pushes

someorg 2026-08-04  (case-mcpx)
Window: 2026-08-04 00:00 - 2026-08-04 07:00 GMT+9

- [overwritten] a/b@main   6 commits gone
- [replaced]    a/c@dev    no common ancestor with the old history
- [unverified]  a/e@main   the commit has been garbage collected

Total: 6 commits gone
1 branch was replaced outright, and a restore cannot bring it back
1 push could not be verified, which does not mean it did not happen

All eleven are read-only

Restore isn't exposed as a tool, and not because it was hard to add.

The agent reads text an attacker wrote. That text can contain "ignore your previous instructions", and a write tool sitting open is the path in.

With reads only, the worst case is a wrong explanation. The repository stays untouched. A person pulls the restore trigger, in the app.

Get started →