Back to Home

How to Connect ChatGPT to OpenClaw Without API Keys

On November 17 at 21:49 UTC, OpenClaw took the beta tag off Sign in with ChatGPT. The authorization route now covers ChatGPT Business workspaces, names GPT-6.2 Astra and Sol as supported models, and stands as a real third option next to Codex login and a plain OpenAI Platform API key. That sounds like a footnote. It changes which subscription pays for your agent, which tools it can reach, and where the failure modes live.

OpenAI's November 10 developer note had already placed OpenClaw beside Notion, v0, Amp, Kilo Code, Warp, Hermes and OpenCode in its Business-plan rollout. The beta exit is OpenClaw's own milestone inside that wider push. If you have been waiting for a reason to stop pasting API keys into a config file, this is it. If you assumed the plan now covers everything ChatGPT does, keep reading, because it does not.

Step 1: Pick the credential that matches the job

OpenClaw uses one provider id, openai, for all three routes, so the model refs stay identical. What changes is the identity behind the request and the services you can actually reach. Choosing badly is the usual reason people call the integration broken.

  • Sign in with ChatGPT: your ChatGPT account and workspace authorize OpenClaw as its own app. Eligible Responses API models run on your included plan usage. Best when you want app-specific consent and plan-based billing.
  • Codex login: your account uses the Codex product, with a one-time device code that works on remote machines. Best when you need the Codex allowance and its route to OpenAI-hosted plugins with connector scopes.
  • API key: a Platform project key billed per use. Best when you need image generation, transcription, speech or memory embeddings from the same credential.

The trap is capability creep. Sign in with ChatGPT does not support OpenAI-hosted plugins, connected apps, hosted MCP tools, tool search or hosted image generation. It also does not authorize OpenClaw's image_generate tool, audio transcription, speech synthesis or memory embeddings. Text, images and files work as inputs when the selected model accepts them, but that is not the same as Files API access. Your local OpenClaw tools and locally configured plugins keep their own credentials, so a Slack or GitHub integration set up inside OpenClaw still works.

Step 2: Run the login and approve token sharing

Run this on the machine that actually hosts your OpenClaw install:

openclaw models auth login --provider openai --method siwc

During the browser step you must approve token sharing. Grant identity permissions alone and OpenClaw will save the account, then refuse to run inference until you enable sharing or pick another credential. This is the most common half-configured state: a green checkmark with an agent that cannot answer.

The browser returns to http://localhost:PORT/auth/callback, and OpenClaw picks a free port for every sign-in, so nothing else on that port can block you. If the browser lives on another machine, copy the port from the sign-in link and forward it to OpenClaw's IPv4 loopback, for example ssh -N -L PORT:127.0.0.1:PORT user@gateway-host, then open the link there. The port changes every attempt, so yesterday's tunnel will not carry over.

To attach the account to a person rather than an agent, use openclaw models accounts login openai --method siwc instead. Agent credentials and personal accounts are separate: the personal default applies to new chats, while existing chats keep whatever account they started with. When saving an agent credential, --agent targets a configured agent, --profile-id names the profile, and --set-default also switches to the provider's recommended model. Leaving that last flag off preserves your current default, which is usually what you want on a tuned setup.

Step 3: Check the quota and seal the seams

Open a conversation and run /status to confirm which account, endpoint and credential the agent is using, then send a real message. A saved account is not proof of working inference, and a populated model list is not proof of current access, because OpenClaw discovers SIWC models from the selected account.

On the billing side, connected apps consume your existing plan usage. There is no extra allowance for adding OpenClaw. ChatGPT exposes a weekly per-app limit under Settings, Usage, and the Control UI's Manage usage button opens the same page. On Business, a workspace sharing cap sits above the per-app limit, which means an individual member can set a lower ceiling but cannot raise the workspace one.

  • Admins: in the OpenAI Admin Console, open Tenant settings, External access, pick the app, and set Use ChatGPT plan to Allowed.
  • Admins: set the sharing cap under Usage limits, ChatGPT plan, Usage in other services. The percentage applies to each member's included allowance across approved apps.
  • Users: when the workspace cap blocks a request, use Copy request for admin and send the message along rather than retrying in a loop.

Credential handling is quieter than it used to be. OpenClaw refreshes SIWC credentials on use within five minutes of expiry and caps stored JWT expiry at the token's exp claim. If a turn hits token_expired, it refreshes the profile and retries that model once before falling back, so a stale token usually costs you one retry rather than a dead session.

One runtime caveat remains. With the native Codex harness, SIWC needs a managed local process and an isolated agent home. Automatic in-turn compaction works, but manual /compact, remote execution and supervised sessions do not. OpenClaw's own runtime has none of those restrictions, so if you rely on manual compaction, keep the agent on OpenClaw's runtime and treat Sign in with ChatGPT as the model credential only.

The beta exit is an access milestone, not a capacity increase. Spend ten minutes this week verifying which route your agent uses, whether token sharing is actually granted, and what your weekly app limit looks like before a long autonomous job runs into a wall you forgot you built.

Comments

No comments yet. Be the first to share your thoughts!