RookeryOS — Platform Overview

The Operating System Where the LLM is the Kernel and the Browser is the API

What is RookeryOS?

RookeryOS is the operating system layer for AgentWorks Studio 2.0 — a platform that deploys autonomous AI agents as containerized pods on Kubernetes. Each pod is a complete dual-container environment where:

1. The LLM acts as the kernel — Hudson Gateway orchestrates agent logic, tool calls, and memory
2. The browser is the universal API — Instead of integrating 335+ SaaS APIs, agents use a headless browser to interact with any web service
3. Humans authenticate on behalf of agents — When a login wall appears, a human co-pilot steps in via noVNC, and cookies are captured for future sessions

The Problem It Solves

API Integration Hell

Traditional agent platforms require dedicated API connectors for every service:

Problem	Impact
335+ API connectors across 25 industries	Unsustainable maintenance burden
OAuth complexity — every service has different auth flows	Brittle, constantly breaking
Schema drift — APIs change without notice	Agents fail silently
Rate limits & pricing — per-API costs add up	Unpredictable expenses
21 MCP servers for external SaaS	Questionable ROI, high overhead

No Tier Isolation

Free-tier Navigator pods and production pods share the same cluster resources — no resource boundaries, no network isolation, no lifecycle management.

Siloed Browser Automation

Existing browser tools (Playwright, CDP) run inside the agent container, competing for resources and creating security risks.

The RookeryOS Solution

1. Browser as Universal API

Instead of maintaining API connectors, agents use the browser:

Traditional:  Agent → API SDK → External Service (breaks when API changes)
RookeryOS:   Agent → Browser → External Service (works as long as the website works)

The browser is the most stable API in existence. Websites maintain backwards compatibility for decades. When a human can use a website, an RookeryOS agent can too.

2. Human-in-the-Loop Authentication

When an agent encounters a login wall:

1. Login Detector identifies authentication pages via DOM heuristics
2. HITL Controller pauses agent execution
3. WebSocket event alerts the Platform Admin console
4. Human clicks alert → noVNC panel opens → authenticates manually
5. Cookie Vault captures and encrypts the session
6. Agent resumes with authenticated cookies injected

Future visits to the same service use the stored cookies — no human needed again until they expire.

3. Encrypted Cookie Vault

• AES-256-GCM encryption with per-tenant Data Encryption Keys (DEK)
• DEKs stored in GCP Secret Manager
• Cookie capture happens automatically after human authentication
• Cookie injection before every page.goto() call

4. Dual-Container Isolation

Each pod runs two containers with strict separation of concerns:

Container	Role	Technology
Kernel	Agent execution engine	Hudson Gateway (Node.js 22), nginx, s6-overlay
Sidecar (Eye/Hand)	Browser + human view	Chromium, noVNC, Browser-Use (Python)

The kernel never runs browser code. The sidecar never runs agent logic. They communicate via CDP (Chrome DevTools Protocol) on localhost:9222.

5. Navigator Tier Isolation

Free-tier Navigator pods are physically isolated:

Aspect	Production	Navigator
Namespace	Per-agent (agent-blue-heron-XXXX)	Shared navigator
Browser	Yes (dual-container)	No
Cookie Vault	Yes	No
Persistent Storage	Yes (10Gi PVC)	No (ephemeral)
Lifetime	Unlimited	24 hours max
Resources	Tier-based	200m/256Mi strict

Platform Components at a Glance

┌─────────────────────────────────────────────────────────────┐
│                    AgentWorks Studio 2.0                     │
│                                                              │
│  ┌──────────────┐  ┌──────────────┐  ┌──────────────────┐   │
│  │  Marketing   │  │   Console    │  │   API Server     │   │
│  │  (Next.js)   │  │  (React 19)  │  │  (Hono + PG)     │   │
│  │  :3000       │  │  :3000       │  │  :3001           │   │
│  └──────────────┘  └──────┬───────┘  └────────┬─────────┘   │
│                           │                    │              │
│                    ┌──────┴────────────────────┴─────────┐   │
│                    │         Lighthouse Knowledge         │   │
│                    │    (Redis Registry + PG + pgvector)  │   │
│                    └────────────────┬────────────────────┘   │
│                                     │                         │
│  ┌──────────────────────────────────┴──────────────────────┐ │
│  │              Agent Pod (Dual-Container)                  │ │
│  │                                                          │ │
│  │  ┌─────────────────┐         ┌─────────────────────┐    │ │
│  │  │  Kernel          │   CDP   │  Sidecar            │    │ │
│  │  │  Hudson Gateway  │◄───────►│  Chromium + noVNC   │    │ │
│  │  │  nginx           │ :9222   │  Browser-Use        │    │ │
│  │  └─────────────────┘         └─────────────────────┘    │ │
│  └──────────────────────────────────────────────────────────┘ │
│                                                              │
│  ┌──────────────┐  ┌──────────────┐  ┌──────────────────┐   │
│  │  Sonar STT   │  │  Sonar TTS   │  │  Agent           │   │
│  │  (Whisper)   │  │  (Local)     │  │  Provisioner     │   │
│  └──────────────┘  └──────────────┘  └──────────────────┘   │
└─────────────────────────────────────────────────────────────┘

Key Metrics

Metric	Value
API Connectors Replaced	335+
MCP Servers Eliminated	21
Cookie Encryption	AES-256-GCM
Browser Latency	<100ms (CDP localhost)
Voice Latency	<300ms (WebRTC local)
Pod Spin-up	~45s (Kernel) + ~60s (Sidecar)
Navigator TTL	24 hours
Base Image Size	~200MB (hardened Debian)
Monthly Infra Cost	~$33 (Navigator) to ~$337 (Full Platform)

What's Next

• Tech Stack — Every technology in the platform
• Dual-Container Architecture — How the pod works
• Lighthouse Knowledge Layer — The state-driven knowledge system
• Security Model — Encryption, isolation, and access control