counterpoint

counterpoint is a tool that makes two different AI models check each other's work before anything gets shipped inside Claude Code. Instead of you copying a draft from one AI, pasting it into another AI to ask what is wrong with it, and then pasting the feedback back, this project automates that whole back and forth on every turn.

It uses Claude Opus as the model that writes the work, and OpenAI Codex, reached through its own command line tool, as a second model that reviews it. Because the two models come from different companies and were trained differently, they tend to make different mistakes, so one can catch a mistake the other would not notice on its own. The README gives an example where one model wrote a data summary calling a certain number weak, and the other model caught that the number was actually a misleading average, and the real result was strong.

The tool needs both a Claude subscription and a ChatGPT subscription already signed in, since it runs both AI apps you already pay for rather than calling a separate paid API. There is no extra per-token billing, but each turn does take longer, since a full check from the second model can add several minutes.

There are three modes you can switch between: off, which runs Claude Code normally with no second opinion, small, which adds one round of drafting, criticism, and revision, and full, which also has both models plan the work first, compares the two plans, and lets the stronger plan's author do the actual writing while the other reviews it. Disagreements between the two models are settled first by any automated checks like tests or a linter, then by whether the work follows the written instructions, and only as a last resort by one model's judgment call.

The project is written mostly as shell scripts and is designed to run on macOS or Linux. The full README is longer than what was shown.