The Platform
An operating system for autonomous agents
RookeryOS treats each agent as a complete machine: the LLM is the kernel, a real browser is the universal API, and every pod is isolated, observable, and secure by default.
Dual-container architecture
Two containers, one machine
Rather than cramming the agent, the browser, and a VNC server into one bloated image, RookeryOS splits responsibilities into two containers that share a pod network and talk over the local Chrome DevTools Protocol — under 100 milliseconds, no external hops.
- The Kernel
- The agent execution engine — orchestrates model calls, tools, memory, and a terminal sandbox. It never runs browser code.
- The Sidecar
- A headless Chromium with a live human-viewable stream and a DOM-distillation service that turns pages into compact, model-readable content. It never runs agent logic.
Agent engine · tools · memory · terminal
Chromium · noVNC · DOM distillation
Built-in systems
Everything an agent needs to do real work
Browser as universal API
Operate any web app a human can — no per-service connector, no OAuth maze, no schema drift.
Human-in-the-loop auth
Pause on login walls, hand off to a person over a live stream, then capture the session once.
Lighthouse knowledge
Sub-100ms registry and semantic search answer from a source of truth before the model is called.
Sonar voice engine
Local speech-to-text and text-to-speech over WebRTC for sub-second, low-cost voice.
Terminal sandbox
A restricted shell lets the agent pick the efficient path — a quick request or a full browser render.
Structured output
Findings are formatted into clean documents and rendered straight into the operator dashboard.
Agent pod lifecycle
From request to teardown
- 1
Create
A request from the Agent Factory or API defines the agent — its persona, tools, tier, and capabilities.
- 2
Provision
The provisioner generates an isolated namespace, RBAC, network policy, storage, and routing on Kubernetes.
- 3
Bootstrap
A oneshot validator waits for the browser, mounts persistent storage, and restores any saved sessions.
- 4
Run
The kernel serves agent requests — checking knowledge first, driving the browser, and calling the model only when needed.
- 5
Teardown
On delete or expiry, all resources are removed and sessions cleaned up — knowledge is preserved at the tenant level.
Tier isolation
Free-tier work never touches production
Free assessment pods run in a physically separate Kubernetes namespace with strict resource quotas, restricted networking, and no persistent storage.
| Aspect | Production | Navigator (free) |
|---|---|---|
| Namespace | Per-agent, isolated | Shared, restricted |
| Browser sidecar | Yes | No |
| Cookie Vault | Yes | No |
| Persistent storage | Yes | Ephemeral only |
| Lifetime | Unlimited | 24 hours max |
See agentic work in action
RookeryOS powers AgentWorks Studio AI Agent Pods. Reach out for a walkthrough or early access.