Review with a ChatGPT Subscription¶

The chatgpt provider reviews with your ChatGPT Plus or Pro subscription. You sign in once with a code in your browser. There is no API key to create or pay for.

It is for reviews on your own machine: lgtmaybe review, lgtmaybe diagram and lgtmaybe mcp. For CI, use the OpenAI provider with an API key.

lgtmaybe login chatgpt

lgtmaybe prints a URL and a one-time code. Open the URL, sign in to ChatGPT and enter the code. The command waits until you approve it, for up to 15 minutes. It then prints where the token is stored.

The sign-in is the same one the Codex CLI uses. lgtmaybe signs in through litellm, which keeps the token at ~/.config/litellm/chatgpt/auth.json and refreshes it for you. You only sign in again if the token is revoked or stops refreshing.

Never share the code. Device codes are a common phishing target.

Run a Review¶

lgtmaybe review --provider chatgpt --model gpt-5.5

Provider and model are not secrets, so they can live in .lgtmaybe.yml:

provider: chatgpt
model: gpt-5.5

The token stays in litellm's file. lgtmaybe never copies it into its own config.

Use It from a Coding Agent¶

The MCP server takes the same flags:

claude mcp add lgtmaybe -- lgtmaybe mcp --provider chatgpt --model gpt-5.5

Sign in with lgtmaybe login chatgpt before you start the agent. The server never asks you to sign in itself. See Use from a Coding Agent over MCP.

Choosing the Model¶

Pass a model your plan can use in Codex, such as gpt-5.5 or gpt-5.3-codex. Use the name without a prefix. lgtmaybe adds the chatgpt/ route for you.

Reviews count against your plan's Codex usage limits. A review sends one request per lens, several at once. On a large change, a Plus plan can hit its limit part-way through. If it does, set max_concurrency: 2 in .lgtmaybe.yml or review a smaller change.

What Differs from the OpenAI Provider¶

The subscription reaches OpenAI's Codex backend, not the public API. That backend accepts fewer request settings:

The review format is sent as a tool call. lgtmaybe still holds the model to its JSON schema. Nothing changes in what gets posted.
temperature is ignored. The default of 0.0 cannot reach the model. So two runs on the same change can differ, and the run log says so.
max_tokens is ignored. The backend applies its own output limit.

The backend is not a published API. OpenAI can change it without notice, and an lgtmaybe update may be needed when it does.

Troubleshooting¶

"no sign-in was found": run lgtmaybe login chatgpt. The message names the file it looked for. If you set CHATGPT_TOKEN_DIR, check that it points at the same directory you signed in with.

"the ChatGPT sign-in is missing or has expired": the stored token could not be refreshed. Run lgtmaybe login chatgpt again.

A review in CI fails with "no sign-in was found": this is expected. A CI runner has no stored sign-in. Use --provider openai with an OPENAI_API_KEY secret there.