Vercel AI SDK runs coding agents on your Claude or Codex plan
The harness layer now signs nine agents — Claude Code, Codex, Copilot and more — in through their own logins; OAuth stays on the host, out of the sandbox.
Copy markdown
One login, nine harnesses
The AI SDK harness layer now authenticates Claude Code, Cline, Codex, Cursor, fx, GitHub Copilot, Grok Build, OpenCode and Pi through their own subscriptions — no per-provider API keys to wire up. If the user is already signed into the agent, so is your app.
Tokens never touch the sandbox
Credentials stay at the host boundary: OAuth is refreshed server-side, the sandbox runs on placeholder creds, and the real token is injected into outbound requests on the host. You can execute untrusted agent code without leaking anyone's login.
Three modes decide who pays
`direct` uses explicit provider env credentials and falls back to the native subscription; `auto` matches it until you set an AI Gateway key; `ai-gateway` never reads subscriptions. One flag is the difference between billing a user's seat and billing your metered Gateway account.
Build this weekend: an agent product on other people's plans
Drop a real coding agent — Claude Code or Codex in a Vercel Sandbox via HarnessAgent — into your own app and bill the work to each user's Claude, ChatGPT or Copilot plan instead of your API budget. The subscription economy, for agents you embed.