The Hardware Layer for the Agentic Economy
Give your AI Agents the compute they need to evolve. Rent GPUs by the second. Pay permissionlessly with USDC on Base.
Legacy cloud infrastructure is built for humans, not Agents.
- Identity Wall: Agents can't pass KYC or sign up for AWS/GCP.
- Payment Barrier: Credit cards and bank accounts are for humans. Agents have wallets.
- The "Brain" Gap: Local hardware (MacBooks) is too weak for fine-tuning. Agents need high-VRAM GPUs for self-improvement.
- Manual Overhead: Provisioning instances in a dashboard is a human-only task.
- Untapped Supply: Millions of idle 3090s/4090s sit in gaming rigs and workstations.
- Friction: Monetizing a single GPU is too complex for most owners.
- Trust: No secure way to rent out hardware to anonymous software.
The result? Most AI agents are "locked" in their current state, unable to scale their own intelligence because they can't access hardware.
ComputeSwarm is The General-Purpose Compute API for Autonomous Software.
- Permissionless: No signups. Auth is a wallet signature.
- Pay Per Second: No minimums. Stream USDC as the code executes.
- Any Workload: Fine-tune, infer, process data, generate images—send any Docker job.
- API-First: Designed to be integrated into autonomous loops.
- Bee-Simple Setup: The "One-Click Hive" agent turns your GPU into a worker node.
- Streamed Income: Watch USDC flow into your wallet in real-time.
- Sandboxed Security: Two-phase execution ensures jobs are air-gapped and safe.
x402 Protocol enables trustless, per-second micropayments in USDC on Base L2.
No middleman. No minimum commitment. No trust required.
This project is built for the x402 hackathon and demonstrates real-world use of the x402 payment protocol:
+-------------+ Submit Job +-------------+ x402 Payment +-------------+
| Buyer | ---------------> | Marketplace | <----------------> | Seller |
| (has $) | | (queue) | | (has GPU) |
+-------------+ +-------------+ +-------------+
| |
| USDC on Base |
+--------------------------------------------------------------------+
How x402 is used:
- Buyer submits job with max price - escrowed in smart contract
- Seller claims job - validates buyer has funds via x402
- Job executes - per-second billing tracked
- Job completes - x402 settles USDC payment to seller
| Feature | Description |
|---|---|
| x402 Payments | Trustless USDC micropayments on Base L2 |
| GPU Auto-Detection | NVIDIA CUDA, Apple Silicon MPS, AMD ROCm |
| Multi-GPU Support | Use all your GPUs (2x RTX 4090? No problem) |
| Docker Sandboxing | Secure execution with network isolation |
| Job Templates | Pre-built templates for PyTorch, HuggingFace, LoRA |
| Model Caching | Persistent HuggingFace/PyTorch cache across jobs |
| Real-time Earnings | Track your earnings as a seller |
| Cost Estimation | Know costs before submitting |
Powerful AI compute, finally affordable for everyone:
# Instead of $200/month for cloud GPU access...
# Train your AI model for $0.50 on ComputeSwarm- Students: Experiment with cutting-edge AI without department budgets
- Researchers: Run experiments that would cost thousands in cloud fees
- Indie developers: Train custom AI models for your apps and games
- Small businesses: Add AI capabilities without massive infrastructure costs
- Hobbyists: Turn your ideas into reality with real GPU power
Better utilization of your existing compute infrastructure:
# Your idle data center GPUs can now earn revenue
# Instead of wasting electricity, generate income- Cloud providers: Offer cheaper GPU access through decentralized network
- Tech companies: Monetize overnight idle time in development workstations
- Gaming companies: Utilize player GPUs during off-peak hours
- Research institutions: Share compute across departments and universities
- Hardware manufacturers: Create new revenue streams for GPU owners
The market opportunity: Millions of GPUs worldwide sit idle 80% of the time. That's billions in wasted compute capacity that can now be unlocked.
- GPU Market Size: $50B annually (gaming, workstations, data centers)
- Idle Capacity: 80% of GPUs unused globally = $40B in wasted hardware
- Current Cloud GPU Pricing: $2-4/hour = $17,520-35,040/month
- ComputeSwarm Pricing: $0.50-2.00/hour = 75-90% cost reduction
- Addressable Market: Anyone currently paying for cloud GPUs (developers, researchers, companies)
We're not competing with cloud providers. We're utilizing the compute they can't access.
- Python 3.10+
- Node.js 20+ (for frontend)
- Docker (for sandboxed execution)
- NVIDIA GPU with CUDA (or Apple Silicon)
git clone https://github.com/Archdiner/compute-swarm.git
cd compute-swarm
python -m venv venv
source venv/bin/activate
pip install -r requirements.txtcp .env.example .env
# Edit .env with your settings:
# SUPABASE_URL=your-project-url
# SUPABASE_ANON_KEY=your-anon-key
# SELLER_PRIVATE_KEY=0x... (for receiving payments)
# BUYER_PRIVATE_KEY=0x... (for sending payments)# Terminal 1: Start marketplace server
python -m src.marketplace.server
# Terminal 2: Start frontend web UI
cd frontend
# Create .env.local (already created if following setup)
npm install
npm run dev
# Note: Users need MetaMask browser extension to connect wallets
# Terminal 3: Start seller agent (on GPU machine)
python -m src.seller.agentThen visit http://localhost:3000 in your browser.
# Terminal 1: Start marketplace server
python -m src.marketplace.server
# Terminal 2: Start seller agent (on GPU machine)
python -m src.seller.agent
# Terminal 3: Submit jobs as buyer
python -m src.buyer.cli# Interactive mode
python -m src.buyer.cli
> templates # See available job templates
> template # Submit using a template
Select: pytorch_train
epochs: 10
batch_size: 64
...
> status abc-123 # Check job status
Status: COMPLETED
Output: Training complete! Accuracy: 94.2%
Cost: $0.0847 USDC# One-click setup
./scripts/setup_seller.sh
# Start earning
python -m src.seller.agent
# Output:
# ============================================================
# ComputeSwarm Seller Agent
# ============================================================
# Node ID: node_a1b2c3d4e5f6
# GPU: NVIDIA GeForce RTX 4090
# Price: $2.00/hr
# Status: Available
# ------------------------------------------------------------
# Session Earnings: $0.0000 USDC
# Jobs Completed: 0
# ============================================================compute-swarm/
├── frontend/ # React web UI with Privy embedded wallets
│ ├── src/
│ │ ├── components/ # React components
│ │ ├── pages/ # Buyer/Seller views
│ │ ├── services/ # API client, Privy config
│ │ └── hooks/ # Custom React hooks
│ └── package.json
├── src/
│ ├── marketplace/ # FastAPI server (job queue, matching, payments)
│ ├── seller/ # Seller agent (GPU detection, job execution)
│ ├── buyer/ # Buyer CLI (job submission, templates)
│ ├── execution/ # Docker sandboxing, GPU passthrough
│ ├── payments/ # x402 USDC payment processing
│ ├── templates/ # Pre-built ML job templates
│ └── config.py # Configuration management
├── docker/ # Docker images for sandboxed execution
├── scripts/ # Setup and helper scripts
└── tests/ # Test suite
Tech Stack:
- Frontend: React, TypeScript, Vite, MetaMask (wallet connection), Tailwind CSS
- Backend: FastAPI, Python 3.11
- Database: Supabase (PostgreSQL)
- Payments: x402 SDK, Web3.py, USDC on Base
- Compute: PyTorch, Docker, NVIDIA Container Toolkit
- Authentication: MetaMask wallet connection
Pre-built templates for common ML tasks:
| Template | Description | GPU Required |
|---|---|---|
pytorch_train |
Train PyTorch models | Yes |
huggingface_inference |
Run HuggingFace models | Yes |
lora_finetune |
LoRA fine-tuning | Yes |
image_classification |
Classify images with pretrained models | Yes |
gpu_benchmark |
Benchmark GPU performance | Yes |
| Endpoint | Method | Description |
|---|---|---|
/api/v1/jobs/submit |
POST | Submit a job to the queue |
/api/v1/jobs/estimate |
POST | Estimate job cost |
/api/v1/jobs/{id} |
GET | Get job status |
/api/v1/nodes |
GET | List available GPU nodes |
/api/v1/stats |
GET | Marketplace statistics |
/api/v1/sellers/{addr}/earnings |
GET | Seller earnings dashboard |
Full API docs at /docs when running.
- x402 USDC payment integration
- Multi-GPU support
- Docker sandboxing
- Job templates
- Seller earnings dashboard
- Web dashboard for buyers and sellers
- Privy embedded wallet integration
- Reputation system
- Spot pricing / auctions
- Model marketplace integration
ComputeSwarm is building the decentralized infrastructure for the AI economy.
By connecting idle GPUs with compute demand through trustless micropayments, we're:
- Democratizing AI development - Making powerful compute accessible to everyone
- Optimizing global compute efficiency - Turning waste into wealth
- Accelerating AI innovation - Removing cost barriers to experimentation
- Building resilient infrastructure - Decentralized compute that can't be controlled by any single entity
This isn't just a marketplace. It's a fundamental shift in how compute resources are allocated globally.
This project demonstrates how x402 can enable:
- Micropayments for compute - Pay per second, not per hour
- Trustless transactions - No escrow service needed
- The agentic economy - AI agents can autonomously purchase compute
MIT License - See LICENSE for details.
Made for the x402 Hackathon