What is Orca?
Orca is a hosted, multi-tenant platform for AI agents. You define a reusable agent once, its runtime, model, instructions, tools, skills, and memory, then run it, orchestrate it with other agents, and publish it as a chat endpoint your own apps can call. Every run streams live, every workspace is isolated per tenant, and usage, spend caps, and observability come standard. Under the hood Orca separates a stateless control plane (the Conductor) from stateful session management (Runners), so the platform scales horizontally to coordinate many concurrent agents across multiple LLM providers. You never operate that split on the hosted product, it’s what keeps your agents fast and reliable.Dashboard and API
Create your account, build your first agent, and run it, in a few minutes.
Core Concepts
Agents, sessions, runs, pods, workflows, skills, memory, and more.
SDKs
Typed TypeScript, Python, and Go clients over the same API.
CLI
The
orca command line: agents, runs, and everything else from your terminal or CI.API Reference
Every Conductor route and the chat gateway.
Getting started
- Sign up at www.orcapods.ai and create (or join) an organization.
- Use the dashboard at app.orcapods.ai to build and run agents visually.
- Call the API / SDKs with an
ao_tenant key (base URLhttps://api.orcapods.ai) for application code and automation. - Use the CLI (
curl -fsSL https://orcapods.ai/install.sh | sh, thenorca login) for terminal work, scripts, and CI. See the CLI overview.
Orca is also self-hostable: source access is available to design partners and licensed
self-hosters (the repository is currently private; email support@okik.io for access), and it
runs on Docker Compose or Kubernetes. See Local development and the
deployment guides. The rest of this documentation leads with the hosted product; self-hosting
notes are called out where they differ.
What you can build
Reusable agents
Define an agent as a profile: a runtime, a model, a system prompt, and the tools, skills,
and MCP servers it can use. Launch it anywhere, as many times as you like.
Multi-provider
Run on Claude, OpenAI, Google, Groq, or OpenRouter, chosen per agent, not per deployment.
Live streaming
Every run streams progress, tool calls, and token usage over Server-Sent Events, with replay,
in the dashboard and via the API.
Multi-agent workflows
Compose agents into pods and DAG workflows; a deterministic engine schedules and runs
nodes in parallel.
Skills & memory
Give agents packaged Skills and a per-agent Memory Bank that persists useful knowledge
across sessions.
Tools & MCP
Built-in platform tools plus any external Model Context Protocol server, including SaaS
apps connected through OAuth.
Publish as chat
Publish an agent behind a stable chat endpoint and mint scoped API keys for your app.
Isolated files & storage
Each session gets an isolated virtual filesystem; a shared artefact store holds agent
outputs.
Credits, caps & RBAC
Prepaid credits, per-tenant spend caps, usage metering, and role-based access control
keep teams safe by default.
How it works
- A dashboard action, SDK call, or
POST /api/runsrequest names an agent profile and a prompt. - The Conductor places a session on a capable Runner.
- The Runner builds a session-scoped toolkit and MCP endpoint, then hands the run to its agent-worker sidecar, which drives the chosen LLM.
- Run events stream back through the Conductor to every subscriber in real time.
Supported runtimes
Each agent picks a runtime, which selects how its model is driven:general is accepted as a legacy alias for the vercel runtime.
Models are named provider:model, e.g. anthropic:claude-haiku-4-5,
anthropic:claude-sonnet-4-6, openai:gpt-5.2, or google:gemini-2.5-pro. Availability and
pricing follow the platform’s live rate feed.
Where to next
Build your first agent
The fastest path from sign-up to a running agent.
Agent profiles
How agents are defined and configured.
Publish an agent
Expose an agent as a chat endpoint for your app.