Animated Octomind icon — purple pixel-art octopus

Does ChatGPT Plus include API access? No API key, and no API credits: OpenAI bills API-platform usage separately from a ChatGPT subscription. There's a second route, though. With OpenAI's Sign in with ChatGPT, a supported open-source app can request permission to count eligible requests against your ChatGPT plan. Octomind 0.56.0 adds that sign-in to its local CLI.

Octomind 0.56.0 lets you sign in to your ChatGPT plan, then select an allowed model with chatgpt:<model>. It uses OpenAI's official account authorization flow; you don't paste an API key into Octomind. Plan eligibility and available models depend on the account and the permission you grant. It's out for the open-source CLI today.

Does ChatGPT Plus include API access?

No: ChatGPT Plus doesn't have API access of its own; it gives you no OpenAI API key and no API-platform credits. OpenAI's billing guide says API usage is billed separately from your ChatGPT subscription. OpenAI's plan-sharing help page covers the other route — using your plan in a supported tool instead of API billing (checked October 2026).

RouteSign in withWho pays for model useLimits and where it runsWhat it cannot do
OpenAI Codex CLIChatGPT account, or another available sign-in methodThe Codex usage included in your plan, when you sign in with ChatGPTRuns locally in your repositoryDrives Codex only; gives you no key for other tools
Octomind 0.56 chatgpt:ChatGPT account through Sign in with ChatGPTYour Plus or Pro plan, if you grant plan usage; it counts toward the same ChatGPT Work and Codex usageLocally: the sign-in browser must run on the machine running Octomind. You can cap an app at a percentage of your weekly usageCreates no API key; restricted Responses API; not in Octomind Cloud
OpenAI APIAPI key, usually OPENAI_API_KEYYour API billing account, pay per useAnywhere you can keep a secret: servers, CI, remote boxesYour ChatGPT plan pays none of it

According to OpenAI's help page, only Plus and Pro can use their plan this way, and the tool has to support it. OpenAI says those requests "count toward the ChatGPT Work and Codex usage included in your plan," so Octomind and Codex draw from the same allowance. OpenAI's September 28 developer guide describes eligible Plus and Pro users trying a local open-source tool "without creating or configuring an API key." The Codex CLI guide documents Codex's own ChatGPT sign-in. Checked October 2026; plan rules can change.

Choose the ChatGPT plan route if you work in an interactive local tool, you have Plus or Pro, and the route's limits (below) don't get in your way. Choose an API key for a server, CI job or remote box, for pay-per-use accounting you control, or when you need request fields the plan route rejects. If you hit a limit, Settings → Usage in ChatGPT shows your usage and each app's cap.

How do I use ChatGPT in a terminal with Octomind?

To use ChatGPT in the terminal, you need Octomind 0.56.0 or later on a machine where you can open a browser. Over SSH, a browser on your laptop can't finish the sign-in: OpenAI redirects to a loopback address and a random port on the machine running Octomind.

Run the login command:

bash
octomind login chatgpt

Approve access in the browser on that machine. Octomind suggests "Octomind" as the app name on OpenAI's consent page. The CLI then prints the account email and a models row with the model slugs your plan allows:

[AUTHOR: paste the models row printed by octomind login chatgpt --force]

Use one of the displayed slugs in a command or config:

bash
octomind run --model chatgpt:<model>

Or make it the default in your config's [model] table:

toml
[model]
name = "chatgpt:<model>"

Replace <model> with a slug from your own login output.

If you're already signed in, Octomind prints the account and tells you to run octomind login chatgpt --force to sign in again. To print the URL instead of opening a browser:

bash
octomind login chatgpt --no-browser

Paste it into a browser on the same machine; a browser anywhere else can't reach the redirect.

To stop using your plan, set [model] name back to your previous provider's model. To sign out completely, disconnect Octomind in ChatGPT settings and delete octolib/chatgpt/auth.json from your config directory: ~/Library/Application Support on macOS, ~/.config on Linux, AppData\Roaming on Windows (dirs crate docs).

What does ChatGPT OAuth authorize?

The ChatGPT OAuth flow is authorization code with PKCE, as described in OpenAI's open-source sign-in guide. Octomind opens OpenAI's consent page, receives the callback on 127.0.0.1, then stores credentials in your operating system's config directory with owner-only access. Access tokens refresh automatically before expiry.

The permission that matters for model usage is chatgpt.tokens.use.direct; if you decline it, Octomind reports ChatGPT plan usage was not granted. Run octomind login chatgpt --force and approve plan usage if you meant to enable it. Identity sign-in alone isn't enough to run chatgpt: requests.

This is a route to eligible Responses API requests, not access to your ChatGPT conversation history. OpenAI's guide treats sign-in and plan usage as two separate permissions and says neither grants access to your conversations. Check the requested permissions before you approve.

What are the ChatGPT plan limits in Octomind?

There's no universal quota number to quote. OpenAI lets you cap each app at a percentage of your overall weekly ChatGPT usage. That cap is carved out of your plan allowance, not added to it, so an app can hit its own cap while the plan still has room (OpenAI's errors guide). Check the live numbers in Settings → Usage.

In the October 2026 preview, OpenAI requires streamed requests with store: false and rejects fields including temperature and top_p. It doesn't provide persistent server-side conversation state using previous_response_id; the client has to send the context itself.

Some Responses tools, including Image Generation, File Search, Code Interpreter, and hosted connectors, are unsupported on this route, and so are audio and video input. Octomind sends no temperature or top_p for chatgpt: models.

If a request comes back with a usage-limit error (subscription_sharing_usage_limit_exceeded), pause and check Settings → Usage. OpenAI says not to assume the whole plan is empty, since an app-specific limit may be what you hit. If you need a rejected field, an always-on credential, or an unattended remote process, use an API key or another provider; API use is billed separately.

One number can look confusing in Octomind's cost report: chatgpt: calls show the API-equivalent per-token price for the model. It estimates what those tokens would cost through the API. It isn't what Octomind charged your ChatGPT plan or an extra API bill.

What this login cannot do

  • Create a ChatGPT Plus API key. It authorizes eligible model requests in this supported tool. It doesn't expose an API credential for scripts or unrelated apps.
  • Use every API capability. OpenAI's preview has unsupported request fields and tools.
  • Work in Octomind Cloud. ChatGPT login is an open-source CLI feature in 0.56.0. It isn't available in Octomind Cloud or the web panel.
  • Get around your plan's limits. When the plan or the app's cap is used up, Octomind reports OpenAI's usage error instead of retrying.

If the Codex workflow fits you better, see how Octomind compares with OpenAI Codex. If you're here because another tool's limits reset too slowly, our guide to Claude Code rate limits covers the ways out.

Also new in 0.56.0

Hosted Octohub login inside ACP editors

Editors that speak ACP can now sign you in to the hosted Octohub. A session opens and slash commands work while you're signed out. A prompt that needs a hosted octohub:* model returns auth_required (-32000), and the editor can then call Octomind's one auth method, octomind-login, which opens the browser device flow.

This only happens when OCTOHUB_API_KEY is unset and OCTOHUB_API_URL doesn't point at a self-hosted hub; self-hosted hubs need no login. See the ACP protocol guide for the editor flow.

Hindsight scores turns without changing them

Hindsight is an opt-in local verifier that scores whether the next user turn is likely to correct the agent, and predicts a correction class. It logs scores, shows a supervisor notice, and adds results to /info and hindsight.jsonl; it doesn't block or alter the agent's response. It ships disabled. To turn it on:

toml
[supervisor.hindsight]
enabled = true
model = "muvon/hindsight"

The Octomind CLI skill

The new octomind Agent Skill teaches a compatible orchestrating agent to delegate work to the Octomind CLI. It includes one-shot runs, named daemon sessions, ACP use, JSONL output, and resume behavior. Install it from the Octomind repository's skill directory and follow the host agent's skill-install instructions.

Learning continues after compaction and one-shot runs

A successful automatic compaction now starts lesson extraction from the pre-compaction snapshot, including a one-message one-shot run. Finished one-shot sessions also extract lessons.

The new learning defaults are min_samples = 10 and max_trial_uses = 40. Existing configs that explicitly set the old 3 and 8 values keep those values; change them directly if you want the new defaults:

toml
[supervisor.learning.evolution]
min_samples = 10
max_trial_uses = 40

Evolution remains off by default. The learning guide explains how its trials and lessons work.

Smaller things

  • Hindsight model recovery. A failed model load is retried at the next turn end; the unavailable notice appears once per process.
  • Windows in the README. The README now names Git Bash or MSYS2 for Windows installs.

Coming soon to Octomind Cloud

Octomind 0.56.0 is coming to Octomind Cloud. Cloud machines pick up a new version the next time they start after the images ship.

Update to Octomind 0.56.0

bash
curl -fsSL https://octomind.run/install.sh | bash -s -- --version 0.56.0
octomind --version

Your config upgrades on first run; existing explicit learning settings stay as written. See the installation guide for other install methods and the full changelog for the release details.

Sources