I spend most of my working life as a GP, translating dense clinical evidence into something people can actually use. The rest of it, increasingly, I spend writing software. Today I am releasing a tool that sits squarely at the intersection of the two: codex-claude-companion, now available as an open-source package on npm and GitHub.
What it is
codex-claude-companion is a companion tool for developers who work with AI coding assistants. It lets one assistant delegate a coding task to another and then have that work independently reviewed before it is accepted — with durable, recoverable sessions and isolated review lanes so nothing is silently lost along the way.
In plain terms: instead of taking a single AI's first answer on trust, the tool builds in a second pair of eyes. One lane does the work; a separate, read-only lane checks it and cites the evidence for what it found. If that sounds familiar, it should — it is how good medicine already works.
Why I built it
In clinical practice, you do not act on a claim because it sounds confident. You act on it because it is backed by evidence, and ideally because a second clinician has looked at the same facts and agreed. Confidence is not correctness.
AI coding assistants are extraordinarily useful, but they have the same weakness a tired doctor has at 2am: they can be fluent and wrong at the same time. The instinct I have built over twenty years — verify before you trust, and separate the person doing the work from the person checking it — turned out to map almost perfectly onto how I wanted my own AI tools to behave. codex-claude-companion is that instinct written as software: delegation, independent verification, and review lanes that report what they actually observed rather than what sounds plausible.
Who it is for
It is for developers who already use AI coding assistants and want a more disciplined workflow around them — one where work is delegated deliberately, sessions survive interruptions and recover cleanly, and changes get an independent review pass before they are accepted. If you have ever wished your AI assistant would check its own work the way a careful reviewer would, this is built for you.
How to try it
The package is open source under the Apache-2.0 licence and requires Node.js 18.18 or newer.
npm install -g codex-claude-companion
Issues, ideas, and pull requests are welcome on the repository.
The doctor who codes
codex-claude-companion is not a one-off. It sits in a wider pattern of open-source work you can check for yourself:
- Merged into Hermes: an authored fix for Codex OAuth token self-healing (#45573 — merged with authorship preserved) and co-authored desktop support for OAuth-gated remote gateways (#37888).
- Merged into oh-my-pi: "Add manual coding-agent GC command".
- Further fixes under review in Hermes, OpenClaw and ByteRover.
- I maintain codex-claude-companion, with more open-source work on my GitHub profile.
People are sometimes surprised that a practising GP also ships developer tools. I am not. The discipline is the same in both worlds — be sceptical of confident claims, demand evidence, and let an independent check catch what a single point of view misses. codex-claude-companion is a small piece of that philosophy made public. I hope it is useful, and I would genuinely like to hear how you use it.