I built ait because the AI coding agents I use daily — Claude Code, Codex CLI, Aider — are fast individually, but no tool lets me put two of them in the same loop. You pick one. It runs. You hope it's right. That's not multi-agent — that's single-agent with chat history.
ait is a local control plane that runs Claude Code, Codex, Aider, Gemini CLI, and Cursor as a team on the same task. One agent investigates, hands the context (what was tried, what failed) to a second agent via AIT_CONTEXT_FILE, a third reviews what was written, and the review gate can block the apply if it finds a critical issue. Everything runs on your machine.
Why this might be interesting:
Multi-model, not multi-tool. Different models catch different bugs. A reviewer agent with a different model and a different prompt catches what self-review misses. The reviewer can block the apply, not just comment.
Cross-agent context handoff. When Claude finishes investigating, Codex picks up actual context, not a paste of the previous chat. Stops the "ramp-up cost twice" problem.
Local control plane. No SaaS. No telemetry. Attempt ledger, review findings, and memory all live in .ait/ next to .git/ in your repo. Works offline. Your prompts and diffs never leave the machine.
Wraps agents you already use. Adapters for Claude Code, Codex CLI, Aider, Gemini CLI, Cursor. Not yet-another-agent.
30-second try, no setup: pipx install ait-vcs && ait demo runs a self-contained 60-second walkthrough with no API keys needed. Shows the full intent → multi-agent attempt → review-blocked-apply flow against a real SQLite ledger.
Alpha. Dogfooded daily on real repos with Claude Code and Codex for the past several weeks. Python 3.14, zero runtime deps, MIT.
Happy to answer questions about the design, why local-first, how the review gate works, how attempts and memory are stored, or why two agents with different prompts catch things one agent with a longer prompt doesn't.
ait is a local control plane that runs Claude Code, Codex, Aider, Gemini CLI, and Cursor as a team on the same task. One agent investigates, hands the context (what was tried, what failed) to a second agent via AIT_CONTEXT_FILE, a third reviews what was written, and the review gate can block the apply if it finds a critical issue. Everything runs on your machine.
Why this might be interesting:
Multi-model, not multi-tool. Different models catch different bugs. A reviewer agent with a different model and a different prompt catches what self-review misses. The reviewer can block the apply, not just comment. Cross-agent context handoff. When Claude finishes investigating, Codex picks up actual context, not a paste of the previous chat. Stops the "ramp-up cost twice" problem. Local control plane. No SaaS. No telemetry. Attempt ledger, review findings, and memory all live in .ait/ next to .git/ in your repo. Works offline. Your prompts and diffs never leave the machine. Wraps agents you already use. Adapters for Claude Code, Codex CLI, Aider, Gemini CLI, Cursor. Not yet-another-agent. 30-second try, no setup: pipx install ait-vcs && ait demo runs a self-contained 60-second walkthrough with no API keys needed. Shows the full intent → multi-agent attempt → review-blocked-apply flow against a real SQLite ledger. Alpha. Dogfooded daily on real repos with Claude Code and Codex for the past several weeks. Python 3.14, zero runtime deps, MIT.
Happy to answer questions about the design, why local-first, how the review gate works, how attempts and memory are stored, or why two agents with different prompts catch things one agent with a longer prompt doesn't.
GitHub: https://github.com/m24927605/ait PyPI: https://pypi.org/project/ait-vcs/ npm: https://www.npmjs.com/package/ait-vcs