One governed path for every model and tool call.

Agents never hold raw provider or tool credentials. The AI Gateway keeps real secrets in a vault, issues short-lived tokens, routes and governs every model and MCP tool call, and writes a complete, attributable audit trail.

Get started Read the docs

Production-ready today

Built on solid ground.

Credential custody

Real provider and tool credentials live in a vault (AWS, Azure, GCP, Kubernetes, or env). Agents get short-lived tokens — never the secret.

Short-lived session tokens

Every session is authorized with a TTL-bound token scoped to its workspace and agent. Revocation is instant; a leaked token expires on its own.

Governed MCP egress

Tool calls route through the gateway with per-server allowlisting and per-call authorization — not a raw outbound connection from the agent.

Full audit trail

Every model and tool request is written to a per-workspace audit stream — a complete, attributable record of AI spend and activity for finance and security.

// Real credentials never reach the agent. Secrets live in a vault; the// session references them by id and the gateway injects them at call time.const session = await fetch(`${ORCA}/v1/sessions`, {  method: "POST",  headers,  body: JSON.stringify({    environment_id: env.id,    agent_id: agent.id,    vault_ids: ["vlt_snowflake", "vlt_stripe"],    mcp_servers: [{ name: "warehouse", url: "https://mcp.internal/warehouse" }],  }),}).then((r) => r.json());// Every model and tool call the agent makes is now routed, authorized,// and written to your workspace audit stream — under a per-agent budget.