Run managed agents on any harness — in your own cloud.

Agent Engine is a managed runtime for production agents. It runs in your own cloud, keeps every event, tool call, file, and credential inside your workspace, and mirrors the Anthropic Managed Agents API — point an Anthropic-compatible client's base URL at Orca and your existing agent code runs unchanged.

Get started Read the docs

Architecture

Inside the Agent Engine

A registry and harness server run your agent; a sandbox isolates its code; the AI Gateway governs every model and tool call. Every transcript and audit log streams out to your own data backbone.

Agent Engine architecture: SDK, CLI and UI connect to the engine's Registry, which pairs with the Harness Server. The Harness Server drives a Sandbox and routes model and tool calls through the AI Gateway, which reaches MCP servers and LLM providers. Transcript Store and Audit Logs stream out to Data Streaming.

Production-ready today

Built on solid ground.

Runs in your cloud

Every event, tool call, file, memory, and credential stays inside your workspace. Nothing is exfiltrated to a vendor.

Anthropic-SDK compatible

Mirrors the managed-agents API surface. Point an Anthropic-compatible client's base URL at your Orca workspace — no rewrite.

Real runtime primitives

Sessions and threads, resource mounts (files, memory stores, Git repos), versioned memory, skills, environments, and vaults — a full control plane, not a wrapper.

Durable and replayable

Sessions resume by replaying their event log. Every step is recoverable and auditable.

// Point any Anthropic-SDK-compatible client at your Orca workspace.// Every event, tool call, and credential stays inside your cloud.const ORCA = "https://acme.orca.cloud";const headers = {  "x-api-key": process.env.ORCA_API_KEY!, // orca_...  "content-type": "application/json",};// Harness and model are configuration, not a rewrite.const agent = await fetch(`${ORCA}/v1/agents`, {  method: "POST",  headers,  body: JSON.stringify({    name: "support-triage",    model: { provider: "anthropic", id: "claude-opus-4-8" },    metadata: { harness: "claude_agent_sdk" },  }),}).then((r) => r.json());// Run it. The full transcript streams back as a governed, replayable log.const session = await fetch(`${ORCA}/v1/sessions`, {  method: "POST",  headers,  body: JSON.stringify({ environment_id: env.id, agent_id: agent.id }),}).then((r) => r.json());