Use Your ChatGPT Plan

If you pay for ChatGPT, Caretz can run on that plan’s usage instead of OpenAI API credits. Run caretz setup, choose Hosted provider, then ChatGPT plan (sign in, no API key). Your browser opens to sign in, and you pick one of your plan’s models as the default.

To add it without changing your default model, sign in from the command line instead:

caretz auth login chatgpt

Opening your browser to sign in with ChatGPT. If it does not open, visit:
  https://auth.openai.com/api/accounts/authorize?...
Signed in to ChatGPT as you@example.com.
Added provider chatgpt to /Users/me/.config/caretz/config.json.
Models on your plan: gpt-6.1-sol, gpt-6-astra, gpt-6.1-terra, ...
Use one with caretz --model chatgpt/gpt-6.1-sol, or /models in a session.

Then pick a model the way you would for any provider:

caretz --model chatgpt/gpt-6.1-sol -p "explain this repository"

Either way you get a chatgpt provider listing the models your plan offers. caretz auth login sets activeModel only if you have none yet. Tokens are kept in credentials.json next to your API keys (see Keep API keys out of config.json ) and refresh on their own, so you sign in again only after logging out or revoking access. caretz auth list shows the account and whether plan usage is allowed.

Sign Out

caretz auth logout chatgpt

This revokes Caretz’s access at OpenAI and deletes the tokens. You can also disconnect Caretz from ChatGPT’s settings; the next request then asks you to run caretz auth login chatgpt again.

When Your Plan Runs Out

When you reach the plan’s usage limit, the turn stops with ChatGPT plan usage limit reached; see ChatGPT Settings > Usage. Caretz doesn’t retry it, because retrying only burns through the wait. Switch to another model with /models, or come back once the limit resets.

Details

Model rules still apply. ChatGPT is a hosted provider, so a directory that keeps cloud models out blocks it. To try it once without changing the rule, pass --allow-model chatgpt/gpt-6.1-sol.

Sign-in on a remote machine. Over SSH, or on Linux with no display, Caretz prints the sign-in URL instead of opening a browser. Open it on any machine and sign in. The browser then fails to load a http://127.0.0.1:1455 page, which is expected: copy the full URL from its address bar and paste it into Caretz. If you’d rather not paste, forward the port first with ssh -L 1455:127.0.0.1:1455 host; the redirect then reaches Caretz directly, and you press Enter to continue.

A second account. caretz auth login chatgpt work signs another account in as provider work, and caretz auth logout work signs it out. A signed-in provider looks like this in config.json, and it can’t set its own baseURL or be marked local:

{
  "provider": {
    "chatgpt": {
      "auth": "chatgpt",
      "models": {
        "gpt-6.1-sol": {}
      }
    }
  }
}

No prompt caching. Plan usage doesn’t cache prompts yet, so each turn sends, and counts, the whole conversation. Compaction keeps long sessions from eating your usage faster than they need to.

Other limits. Plan usage only speaks OpenAI’s Responses API, which Caretz selects for you, and only reaches api.openai.com. The tokens are never sent anywhere else, and the model and sandboxed commands never see them.