Prime Compute is an open marketplace for renting real services: GPU, CPU, and full servers, plus storage, VPN, and workers. Connect a wallet to sign in (RainbowKit + SIWE). Signing in provisions a spend wallet the platform custodies for you: fund it with USDC and it pays for rents automatically as they stream. From there you can rent a service or list one of your own.

Listing means running your own service endpoint and registering it so the broker can route renters to you.

1. Run your service behind a public HTTPS endpoint. For compute this is an x402 seller that charges per unit (use the provider server template in services/ as a starting point); the endpoint is where renters and the meter reach you.
2. Pick a per-unit price. You keep every payment: settlement lands directly in the wallet your endpoint signs with. Prime Compute never holds your earnings.
3. Register the service: on List a server, or for agents POST /api/v1/providers with your alias, endpoint URL, region, price, and type-specific specs.
4. Stay online. The broker only routes to reachable, healthy endpoints; your Compute Score reflects real behavior.

Renting gives you real credentials to a real service.

1. Fund your spend wallet with USDC (the wallet panel shows your address and a faucet).
2. Rent: pick a listing on the marketplace, or for agents POST /api/v1/rents. The broker matches you a provider and the lease goes live.
3. Use what you get. The connect payload depends on the type: SSH host + credentials for compute, a WireGuard profile for VPN, a bucket URL + keys for storage, a submit URL + token for a worker. Connect to the provider directly with those, exactly as you would any real server or service.
4. Pay as it runs. The meter streams USDC per unit from your spend wallet; you only pay for what actually runs, and stopping the lease stops the charges.

Every service is priced per unit and metered as it runs. Time-based services (GPU, CPU, full servers, workers) are priced per second, so we also show an exact per-day figure. Volume services show an honest per-unit rate with an example: VPN is per GB (shown as a cost per 100 GB), storage is per GB-hour (shown as a cost per GB-day). A "charge" is one unit at the listed price; your budget is a count of units, so you always know the ceiling.

An AI broker matches each rent to a provider by reasoning over the live listings against what you asked for. It is soul-driven, not a hardcoded score: its behavior comes from a policy it reasons from, with a deterministic fallback so a model outage never blocks a rent.

Payments settle per unit over x402 on Arc. Each unit is one micro-payment from your custodied spend wallet to the provider's endpoint, recorded as a charge. There is no upfront lump sum and no lock-in: an idle lease accrues nothing, and cancelling stops the stream immediately.

Six service types, each with its own specs and connect payload:

• GPU / CPU / Full Server — time-metered compute; connect over SSH.
• Worker — time-metered job runner; connect via a submit URL + token.
• Storage — GB-hour metered; connect via a bucket URL + access keys.
• VPN — GB metered; connect by loading the returned WireGuard profile.

Autonomous agents are first-class. Register once, then rent and list machine-to-machine.

• POST /api/v1/agents — self-register, returns an API key + a funded-capable wallet.
• GET /api/v1/providers — list the marketplace. POST /api/v1/providers — list your own service.
• POST /api/v1/rents — rent. GET /api/v1/rents/:id — status. POST /api/v1/rents/:id/cancel — stop.
• GET /api/v1/wallet — your wallet address + balance. POST /api/v1/wallet — withdraw USDC to an address.

Over MCP the same actions are tools: discover_providers, rent_compute, rent_status, register_server, wallet_balance, withdraw_funds.