BarrHawk

The Operating System for Agentic Verification & Orchestration

██████╗  █████╗ ██████╗ ██████╗ ██╗  ██╗ █████╗ ██╗    ██╗██╗  ██╗
██╔══██╗██╔══██╗██╔══██╗██╔══██╗██║  ██║██╔══██╗██║    ██║██║ ██╔╝
██████╔╝███████║██████╔╝██████╔╝███████║███████║██║ █╗ ██║█████╔╝
██╔══██╗██╔══██║██╔══██╗██╔══██╗██╔══██║██╔══██║██║███╗██║██╔═██╗
██████╔╝██║  ██║██║  ██║██║  ██║██║  ██║██║  ██║╚███╔███╔╝██║  ██╗
╚═════╝ ╚═╝  ╚═╝╚═╝  ╚═╝╚═╝  ╚═╝╚═╝  ╚═╝╚═╝  ╚═╝ ╚══╝╚══╝ ╚═╝  ╚═╝
                                            v0.5.0 | TRIPARTITE RELEASE

"If your Agent can't test itself, it's just a hallucination."

---

What's New in v0.5.0

• Unified Stack - Single command bun run barrhawk starts everything
• Integrated Dashboard - Live monitoring at http://localhost:7000/dashboard
• Experience System - Self-healing selectors that learn across sessions
• Tool Injection - New tools broadcast to running Igors in real-time
• Lightning Strike - Igor auto-escalates to Claude when stuck
• Hub Mode - Multi-Igor test orchestration with --hub flag

---

Quick Start

# 1. Clone and install
git clone [email protected]:barrhawk/barrhawk_premium_e2e_mcp.git
cd barrhawk_premium_e2e_mcp
bun install

# 2. Start the stack
bun run barrhawk

# 3. Submit a test
curl -X POST http://localhost:7001/plan \
  -H "Content-Type: application/json" \
  -d '{"intent":"take a screenshot of the homepage","url":"https://example.com"}'

Stack Modes

bun run barrhawk           # Full tripartite stack (Bridge, Doctor, Igor, Frank)
bun run barrhawk --minimal # Just Bridge + Igor (lightweight)
bun run barrhawk --hub     # Full stack + Test Orchestration Hub

---

The Tripartite Architecture

Four specialized servers communicating via WebSocket:

Component	Port	Role
Bridge	7000	Message bus, circuit breakers, dashboard
Doctor	7001	Planner, intent parser, experience engine
Igor	7002	Executor, Lightning Strike dual-mode
Frankenstein	7003	Browser control, dynamic tool forge

How It Works

1. You send an intent to Doctor: "test the login flow"
2. Doctor consults experience, generates a plan, selects tools
3. Igor executes the plan step-by-step
4. Frankenstein controls the browser, takes screenshots
5. Experience records what worked for next time

Lightning Strike (Dual-Mode Igor)

Igor starts in "dumb mode" - fast, regex-based execution. When a step fails 3 times, Igor "strikes" - elevating to full Claude reasoning to solve the problem, then powers back down.

---

Endpoints

Doctor (Port 7001)

# Submit a plan
POST /plan
{"intent": "click the login button", "url": "https://example.com"}

# Check plan status
GET /plans
GET /plans/:id

Bridge (Port 7000)

# Health check
GET /health

# Live dashboard
GET /dashboard

Igor (Port 7002)

# Status
GET /status
GET /health

# Current tool bag
GET /toolbag

Frankenstein (Port 7003)

# Health and tools
GET /health
GET /tools

---

Project Structure

/
├── bin/barrhawk.ts          # CLI entry point
├── tripartite/              # The Core Stack
│   ├── bridge/              # Message Bus + Dashboard (7000)
│   ├── doctor/              # Planner + Swarm Logic (7001)
│   ├── igor/                # Executor + Lightning (7002)
│   ├── frankenstein/        # Browser + Tool Forge (7003)
│   └── shared/              # Common utilities
│       ├── experience.ts    # Learning system
│       ├── tool-registry.ts # 120+ tool definitions
│       └── types.ts         # Shared types
├── hub/                     # Test Orchestration (--hub mode)
│   ├── index.ts             # Hub API (7010)
│   ├── coordinator.ts       # Multi-Igor sync (7011)
│   └── igor-db.ts           # Database watcher (7012)
├── packages/                # Feature modules
│   ├── golden/              # Visual regression
│   ├── self-heal/           # Selector strategies
│   └── ...                  # 30+ packages
└── docs/                    # Documentation
    └── specs/               # Component specifications

---

MCP Integration

BarrHawk exposes MCP tools for Claude Code, Cursor, and other AI assistants.

Configure Claude Code

./scripts/generate-mcp-configs.sh

Key Tools

// Natural language automation
frank_execute({ task: "Log into github.com", url: "https://github.com" })

// Parallel testing
frank_swarm_execute({ intent: "Test checkout flow", maxIgors: 4 })

// OS-level control (extensions, dialogs)
frank_os_mouse({ action: "click", x: 100, y: 200 })
frank_os_keyboard({ combo: "ctrl+shift+i" })

// Screenshots
frank_screenshot({ fullPage: true })

---

Experience System

BarrHawk learns from every test run:

• Selectors: Which CSS selectors work for which elements
• Timings: How long actions typically take per site
• Errors: Common error patterns and their fixes
• Sites: Known site patterns with pre-mapped selectors

Data stored in experiencegained/experience.json (configurable via EXPERIENCE_DIR).

---

Development

# Run tests
bun test

# Type check
bun run typecheck

# View logs
tail -f /tmp/tripartite-*.log

---

License

MIT

---

"Trust, but Verify." — BarrHawk