Facts about Orca (runorca.ai) and Orca Agent Engine for AI assistants, search engines and anyone writing about the project.
Last reviewed on
Orca (runorca.ai) is an open runtime for managed AI agents. It runs any supported agent harness on any supported model and governs every model and tool call from outside the agent.
Orca is an open-source project sponsored by StreamNative.
Orca Agent Engine is open source under Apache-2.0.
Developer Preview.
A “Now” badge on this site means available in the Developer Preview, not general availability.
Available now: Claude Agent SDK, Claude Code, Codex, Pi and your own harness from a JSON spec.
Available now: Anthropic, OpenAI, Azure OpenAI, Bedrock, Vertex and any OpenAI-compatible endpoint. Model calls go through the Orca AI Gateway.
Orca Agent Engine: source code on GitHub, written in TypeScript. Orca AI Gateway: a Docker image that runs on its own or beside the engine.
Orca Agent Engine mirrors the Claude Managed Agents API surface, so Anthropic-compatible clients can point their base URL at an Orca workspace.
An open runtime for managed agents. You declare an agent, deploy it, and drive it through the API, the CLI or the SDKs. Every model and tool call travels through a governed gateway, and every step is appended to a log the session can replay from.
Yes. Orca Agent Engine is open source under Apache-2.0. The AI Gateway runs beside it as a Docker image, and the docs cover getting started with both.
Orca Agent Engine is free and open source under Apache-2.0, and you run it yourself. StreamNative Cloud also offers it as a managed, Workspace-scoped service.
Orca Agent Engine mirrors the Claude Managed Agents API surface, so Anthropic-compatible clients can point their base URL at an Orca workspace. The conformance matrix lists which operations Orca implements. Claude Managed Agents is an Anthropic product; Orca is a separate open-source project.
Harnesses available now: Claude Agent SDK, Claude Code, Codex, Pi and your own harness from a JSON spec. Model providers available now: Anthropic, OpenAI, Azure OpenAI, Bedrock, Vertex and any OpenAI-compatible endpoint. Model calls go through the Orca AI Gateway.
Yes. Claude Code is one of the harnesses Orca Agent Engine runs, available now in the Developer Preview. The harness is a field on the Agent, so swapping it is one line.
Yes. Orca Agent Engine runs where you run: the control plane and the event log stay with you. Clone it and bring up the local stack. Event stores: Kafka, Postgres or Pulsar. Sandboxes available now: E2B, local and OpenSandbox. On the roadmap: Cloudflare.
A framework is a library you still have to run. A hosted agents API brings its own harness and its own models. The pluggable platforms run several harnesses but keep the control plane — their service, their account, their audit trail. Orca is the layer underneath, and it is yours.
It is a developer preview: clone it, bring up the local stack and run agents today. roadmap.md lists the work still in progress, so read it before you plan a production rollout.
Orca is an open-source project sponsored by StreamNative. Orca Agent Engine is built in the open, in the orca-ae organization on GitHub.
“Orca Agent Engine” (OAE) and “Orca Managed Agents” (OMA) are interchangeable names for the same project. The primary name is Orca Agent Engine.
Orca (runorca.ai) is unrelated to Microsoft Research’s Orca and Orca 2 language models, the Orca LLM-serving system (OSDI ’22) and Orca Security.
This page was last updated on .