Codex (OpenAI)

Codex, made by OpenAI, is the second AI provider Limn can use. Limn runs the Codex program on your computer, on your own ChatGPT sign-in. It can run every AI feature Limn has. When both Claude and Codex are signed in, Limn reviews each pull request with both at once, and merges the two into one review.

#What you need

• A ChatGPT account whose plan includes Codex.
• Nothing else to install: Limn downloads the Codex program (about 360 MB) the first time it starts.

#Sign in

1. Open Settings, then the AI tab.
2. On the Codex line, select Sign in with ChatGPT.
3. Select Open the ChatGPT sign-in and sign in in your browser. On a computer with no browser, select Use a code instead and enter the code on the page it names.
4. The line reads Signed in with ChatGPT., or names your plan.

From a terminal, limn ai login codex does the same (add --device-auth on a computer with no browser), and limn ai logout codex signs out.

#Your own Codex is left alone

Limn keeps its Codex sign-in in its own folder, ~/.limn/codex. Your own Codex settings and sign-in in ~/.codex are never read or changed, and neither is an OPENAI_API_KEY in your environment.

#How Limn locks Codex down

For reviews, chats and summaries, Limn starts Codex with almost everything switched off:

• No shell, no web access, no images, no plugins and no memories.
• No project instruction files are read.
• Files are read only through Limn’s own read-only tools, limited to the repository under review.
• A safety check approves every tool Codex calls. Limn proves the check is running before each run, and refuses to start if it cannot.

AI Fix and Implement story give Codex a shell, but only inside the sandbox, with no GitHub credentials. See AI agents and the sandbox.

#Using Codex

As the default

Choose Codex under Automatic runs and default on the AI tab, or start Limn with LIMN_AI_PROVIDER=codex. Auto uses Claude when both are signed in.

Beside Claude

With both signed in, reviews run on both by default. A point both raise is shown once, and if one half fails the other half still counts as the review. See Claude and Codex together.

Effort

At Default effort, Codex runs at high on its most thorough model. The same Effort setting applies to both providers. See Effort.

#Limits and your data

ChatGPT reports how much of its weekly limit, and five-hour limit where there is one, your account has used. Limn’s automatic runs stop at their share (50% of the week by default), shown in percentages. See AI subscription usage and spend.

With a ChatGPT sign-in, Codex follows your ChatGPT workspace’s data settings. On a personal plan, check your ChatGPT data controls. See How Anthropic and OpenAI treat your work.