Orca Agent Engine

Orca Agent Engine is a managed runtime for production AI agents. You define agents as versioned registry resources, run them in isolated sandboxes, route model and tool calls through a governed gateway, and keep every event, file, memory, and credential inside your own Workspace. Its core API shares 72 operations with the Claude Managed Agents API, so a client written for those operations can call Orca after you change its base URL. Capabilities beyond that contract are extension groups, and which ones exist depends on the deployment.

Why a runtime for agents

Your organization already knows how to run services in production: identity, isolation, change control, audit. Agents break those assumptions. They hold credentials, execute code they were not shipped with, and call external systems autonomously - usually from a framework process with none of the controls your platform team expects.

Agent Engine closes that gap:

• Agents are resources, not processes. An agent is a versioned registry record - model, system prompt, tools, MCP servers, skills - that you review, pin, and roll back like any other config.
• Every run is a session. Sessions produce an ordered, replayable event stream you can tail live, store, and audit.
• Credentials never reach the agent. Vaults hold the real secrets; the gateway injects short-lived tokens per call and writes an attributable audit trail.
• Code runs in sandboxes. Each session executes inside an isolated environment with declared packages and network policy.

Architecture

• Registry - the control plane. Its core API manages agents, sessions, environments, skills, files, vaults, memory stores, and Git credentials on any deployment; a deployment may add extension groups on top. See Registry.
• Harness server - hosts the agent loop on the harness the agent declares. See Runtime.
• Sandbox - isolated execution for each session's code and tool calls, built from an environment template.
• AI Gateway - vault-aware egress for model providers and MCP servers. Real credentials stay in vaults; calls carry short-lived tokens and land in the audit log. The gateway is a separate Orca product that also runs standalone; Agent Engine routes MCP tool calls through it and can use it as the deployment-wide default for model calls. See Use with Agent Engine for configuration.
• Transcript stream - every session event is appended to a durable stream you can tail, replay, and monitor.

Open at every layer

• Any harness. Each agent declares its harness - Claude Agent SDK, Claude Code, Codex SDK, or Pi SDK - and where its loop runs: in the harness server, or inside the session sandbox. See Choose a harness.
• Any model. The agent names its model in a model field - swap the id, or set a speed tier and reasoning effort, without touching agent code. See Model access.
• Any sandbox. Environments declare packages and networking for pluggable sandbox runtimes. See Environments.

Compatibility with Claude Managed Agents

The registry accepts the anthropic-beta: managed-agents-2026-04-01 header and mirrors the Claude Managed Agents resource shapes, error envelope, and pagination, so Anthropic SDKs and tools can call the shared operations by pointing their base URL at your registry endpoint:

curl "$ORCA_REGISTRY_URL/v1/agents" \
  -H "Authorization: Bearer $ORCA_ACCESS_TOKEN"

Orca also ships its own surfaces for the same API: the ork CLI, the TypeScript SDK, the Python SDK, the Go SDK, and the rendered API reference.

Resources you manage

Deployment

There is one registry API and two ways to run it. Self-host the open-source engine (Apache-2.0) in your own cloud, or use a managed Workspace on StreamNative Cloud. Guides on this site are deployment-neutral: they target the registry endpoint you configure. See Self-hosted deployment.

Next steps