A tool that has 3 AIs review AI-written code

mcp-code-assistant-peers is an automatic review tool for code changes made by . The main idea is that the AI that writes the code should not be the only one checking it. Every code edit starts 3 separate AI reviewers at the same time in background tmux sessions.

Each reviewer starts without the original coding conversation, so it can look at the change from a fresh angle, and it can also build up knowledge of the codebase over time. The review notes are written to `.peer-review` output files and SQLite, then the main AI reads the findings. The tool also sets different safety limits for each reviewer: Claude can only write in a prompt folder, Gemini has blocked, and Codex runs with a separate CODEX_HOME and an empty MCP setup.

The stated motivation is that a 2025 survey found 59% of developers often see AI tools create deployment errors, and Google’s DORA report linked a 90% rise in AI coding adoption with a 9% rise in bug rates.

Key points

  • Each change is checked by 3 AI reviewers in parallel.
  • The reviewers do not see the original coding chat, giving them a cleaner starting point.
  • Findings are saved in files and SQLite so the main AI can read them back.
  • Claude, Gemini, and Codex reviewers each run with different .
  • The tool is built around the concern that wider AI coding use can also raise error and bug risk.
Read original