The problem
AI code review tools comment a lot, are often wrong and remember nothing. Teams end up ignoring them. I wanted a reviewer that posts little, posts right, and gets better every time someone tells it it was wrong.
What Donkey does
- Inline reviews with suggestions: comments land on the exact lines, with one-click
suggestionblocks and a check run on the commit. - Risk score: every finding gets a 0–100 score; only those above the chosen threshold are posted on the PR.
- Verifier pass: a second AI pass re-reads each finding in context and rejects false positives before anything is posted.
- Code graph: Donkey knows who calls a changed function, which signatures changed and what has no tests.
- Memory:
@donkey remember …, dismissed findings and resolved threads become learnings an admin approves. - Specialist agents (security, Laravel, architecture, testing, performance…) and cost tracking by feature, repository and model, with a monthly budget.
Under the hood
A Laravel application installed as a GitHub App, self-hosted with its own API key. Webhooks are processed on queues with Horizon, review progress streams live over Reverb, and an MCP server exposes review knowledge to coding agents such as Claude Code.
Want to talk about it? hi@lucienpuget.fr