Use a subscription, not API keys

Agent steps can run on a subscription you already pay for, as long as the step uses the runtime that plan’s owner allows.

Plan	Step executor	Profile model.provider	Credential
ChatGPT Plus, Pro, Business	opencode	openai-chatgpt	openai-chatgpt-auth, from azhi chatgpt login
Claude Pro, Max	claude-agent-sdk	anthropic	A plan token from claude setup-token
GitHub Copilot	opencode	github-copilot	github-copilot-token, from azhi copilot login

ChatGPT plan

npx azhi chatgpt login     # open the link, enter the code, sign in with your ChatGPT account
npx azhi chatgpt check     # renews the sign-in once and says whether it works

model: {provider: openai-chatgpt, name: default, credential: openai-chatgpt-auth}

• default is the server’s AZHI_CHATGPT_MODEL (gpt-5.5 unless set).
• Azhi renews the sign-in itself when the access token has less than an hour left.
• OpenAI replaces the refresh token on every renewal. Give Azhi its own sign-in. Do not paste Azhi’s into your own OpenCode or Codex, or theirs into Azhi, or each cancels the other.
• Cost is recorded as 0 (reported, “ChatGPT plan”). Token counts are still recorded. When the plan limit is reached the step fails and says so.

GitHub Copilot

npx azhi copilot login     # or import an existing sign-in
npx azhi copilot check     # asks GitHub whether the saved sign-in works
npx azhi copilot quota     # your AI-credit allowance, as VS Code shows it

model: {provider: github-copilot, name: default}

• Steps count in AI credits against the monthly pool.
• Model names such as claude-opus-5.5 or gpt-5.6-terra must exist on your plan, or the step fails at start. Use default to fall back to one model.
• The PR review workflow runs on this.

Claude plan

claude setup-token                                    # on any machine with Claude Code; prints sk-ant-oat...
npx azhi secret set claude-plan-token --value <token>

# profiles/<name>@1.yaml
model: {provider: anthropic, name: default, credential: claude-plan-token}

# workflow.yaml, on the agent node
executor: claude-agent-sdk

• Anthropic’s terms do not allow a Claude plan in third-party tools such as OpenCode, but do allow it through the Agent SDK, for your own use.
• Azhi refuses a plan token on model-agent and opencode steps.
• workspace checkouts and harness.opencode setups are OpenCode-only, so the PR review reviewers cannot move to a Claude plan.
• Usage counts against the plan’s limits. Cost is unavailable unless you set pricing in the profile.

Your plan, your terms

This is for your own Azhi with your own plan. A hosted Azhi serving other people must use API keys. Check each provider’s current terms before relying on any of this.

You can also sign in from Mission Control: Secrets › Sign in with ChatGPT / GitHub Copilot. azhi signin list shows every sign-in the workspace has, and how each was supplied.