Keys and accounts

Which AI providers your agents use, and how your workspace pays for them.

A customer workspace's agents run on its own provider API keys, saved by an administrator under Settings → Providers & keys: Anthropic (Claude), OpenAI (Codex), Google (Gemini), DeepSeek (DeepSeek and OpenCode), xAI (Grok), Meta (Muse), and Mistral (Vibe). Kimi (Moonshot) runs on a Kimi Code membership instead of a key: an administrator signs it in under Settings → Providers & keys by approving a code in their own browser, and the card there shows whether it is signed in. A key is tested when it is saved. A seat runs only on a provider your workspace has a key for, and each turn is billed to your key, never to another workspace's.

Other model providers can be connected under Settings → Providers & keys → Model providers. The catalog includes direct APIs, cloud platforms, model gateways, local Ollama, and a custom OpenAI-compatible server. A custom server must be reachable over public HTTPS (private, loopback and link-local addresses are refused). A connected provider seats a chat participant; pick its model as <connection>/<model> in the session's model bar. A custom gateway that answers POST /responses can also be marked for Codex tools.

Settings → Accounts shows the provider accounts and their remaining quota that the workspace's agents run on; every member can see it.

Turn and spend limits per person and for the workspace are under Settings → Limits (see Security).