🐒 Monkey Troop

Decentralized AI Compute Grid

Monkey Troop is a FOSS (MIT Licensed) peer-to-peer network that democratizes access to AI inference. Users donate idle GPU time to run LLM inference for others in exchange for time-based credits, similar to folding@home but for AI.

🌟 Vision

Enable anyone to:

• Donate idle GPU compute when not in use locally
• Earn credits based on actual hardware performance (time-based, normalized)
• Use those credits to access high-performance GPUs when needed
• Trust the network through cryptographic verification and proof-of-hardware

Key Features

• Secure P2P Mesh: Direct node-to-node connections via Tailscale/Headscale (WireGuard)
• OpenAI Compatible: Drop-in replacement for any tool using OpenAI API
• Fair Economy: Time-based credits with hardware multipliers (RTX 4090 = 4x, etc.)
• Proof-of-Hardware: Cryptographic benchmarking prevents hardware spoofing
• Trusted Clusters: Create private networks with friends/teams
• Public Commons: Join the global network at troop.100monkeys.ai
• Multi-Engine: Supports Ollama, LM Studio, vLLM, and more

🏗️ Architecture

┌─────────────┐         ┌──────────────────┐         ┌─────────────┐
│   Client    │◄───────►│   Coordinator    │◄───────►│   Worker    │
│  (Sidecar)  │ Tickets │ (troop.monkey.ai)│Discovery│   (Agent)   │
│             │         │                  │         │             │
│ localhost:  │         │ - Redis Registry │         │ - Ollama    │
│   9000      │         │ - PostgreSQL     │         │ - Tailscale │
└─────────────┘         │ - JWT Auth       │         │ - GPU       │
      │                 └──────────────────┘         └─────────────┘
      │                                                      ▲
      │                 Direct P2P Connection               │
      └──────────────────────────────────────────────────────┘
              (Encrypted via Tailscale WireGuard)

Components

1. Coordinator (Python/FastAPI): Discovery, authentication, proof-of-hardware verification
2. Worker (Rust): GPU monitoring, heartbeat broadcasting, JWT verification proxy
3. Client (Rust): Local OpenAI-compatible API proxy for seamless integration
4. Shared (Rust): Common data structures and types

🚀 Quick Start

There are two installation paths depending on your role:

👥 End Users: Join a Network

Install the worker (to donate GPU) or client (to use GPU) to join an existing network.

# Install worker/client binaries
curl -fsSL https://raw.githubusercontent.com/100monkeys-ai/monkey-troop/main/bin/install.sh | bash

# Join the public network
export COORDINATOR_URL="https://troop.100monkeys.ai/api"
tailscale up --login-server=https://troop.100monkeys.ai/vpn --authkey=<provided-key>

# Start donating compute
monkey-troop-worker

# OR use the network (in another terminal)
monkey-troop-client
# Point your AI tool to: http://localhost:9000/v1

🏢 Network Operators: Deploy a Coordinator Hub

Deploy your own coordinator with Headscale VPN for a private network.

# Clone repository on your VPS
git clone https://github.com/monkeytroop/monkey-troop.git
cd monkey-troop

# Run automated installer (interactive)
./bin/install-coordinator.sh

# OR with command-line flags
./bin/install-coordinator.sh \
  --domain troop.example.com \
  --email admin@example.com \
  --routing-mode path \
  --enable-backups

See DEPLOYMENT.md for detailed documentation.

---

Self-Host a Private Cluster (Manual)

For advanced users who want full control, see DEPLOYMENT.md for manual Headscale setup instructions.

🛠️ Development

Prerequisites

• Rust 1.75+ (for worker/client)
• Python 3.11+ (for coordinator)
• Docker & Docker Compose
• PostgreSQL 15+
• Redis 7+

Build from Source

# Clone the repository
git clone https://github.com/monkeytroop/monkey-troop.git
cd monkey-troop

# Build Rust components
cargo build --release

# Set up Python environment
cd coordinator
python -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate
pip install -r requirements.txt

Run Locally

# Start coordinator stack
docker-compose -f docker/docker-compose.coordinator.yml up -d

# Run worker (requires GPU)
cargo run --bin monkey-troop-worker

# Run client
cargo run --bin monkey-troop-client

Using Streaming

Enable streaming responses for real-time token generation:

import requests

response = requests.post(
    "http://localhost:3000/v1/chat/completions",
    json={
        "model": "llama3:8b",
        "messages": [{"role": "user", "content": "Write a story"}],
        "stream": True
    },
    stream=True
)

for chunk in response.iter_lines():
    if chunk:
        print(chunk.decode('utf-8'))

Multi-Engine Support

Monkey Troop automatically detects and supports multiple inference engines:

• vLLM (highest priority - fastest inference)
• Ollama (versatile, easy setup)
• LM Studio (GUI-based management)

Workers detect all available engines at startup and route requests intelligently based on model availability. vLLM models are prioritized for performance.

Setup vLLM (optional):

# Install vLLM
pip install vllm

# Start vLLM server
vllm serve meta-llama/Llama-3-8B --port 8000

# Or use custom host
export VLLM_HOST=http://localhost:8000

Configure model refresh (optional):

# Check for new models every 5 minutes (default: 3 minutes)
export MODEL_REFRESH_INTERVAL=300

📖 Documentation

• docs/DEPLOYMENT.md - Deploy your own Headscale coordinator
• CONTRIBUTING.md - Development setup and guidelines
• docs/PROJECT_STRUCTURE.md - Project architecture details
• docs/TESTING_GUIDE.md - Testing instructions

🤝 Contributing

Monkey Troop is fully open source (MIT License). Contributions are welcome!

See CONTRIBUTING.md for development setup.

📜 License

MIT License - Copyright (c) 2026 Monkey Troop Contributors

🙏 Acknowledgments