Skip to main content

​
What is the LLM gateway?

Claworc includes a built-in LLM gateway — an internal proxy that sits between your agent instances and LLM provider APIs (Anthropic, OpenAI, Ollama, and more). Diagram showing instances sending requests via virtual keys to the LLM Gateway inside the control plane, which forwards them using real API keys to Anthropic, Google, Ollama, and OpenAI Instead of giving each OpenClaw agent a real API key, the control plane issues every instance a virtual key — a unique, per-instance, per-provider credential. OpenClaw uses this virtual key to call the gateway. The gateway validates the key, looks up the real provider credential, and forwards the request. Your real API keys never leave the control plane. See Configuration to add providers and manage API keys, or Assign models to enable providers on a specific instance.

​
Security properties

  • Credentials stay out of containers. Even if an agent container is compromised, there are no real API keys inside it to steal. A container breakout or log leak cannot expose your provider credentials.
  • Encrypted at rest. All provider API keys — both global settings and per-instance overrides — are stored Fernet-encrypted in the database. The encryption key is auto-generated on first run and stored separately.
  • Network isolation. The LLM gateway is never exposed on a public port. It is only reachable by OpenClaw instances.
  • Per-instance revocation. Virtual keys are one-to-one: one key per instance per provider. You can remove a provider from one instance, or delete a provider entirely, without affecting any other instance or rotating your real provider key.
  • Audit trail. Every request through the gateway is logged with timing, token counts, status codes, and error details.