Rent the machine.Pay the machine.
Providers run a daemon on an idle GPU. Renters — people at a terminal, or autonomous agents — pay that node directly, by the second, for a container on it and get back whatever they do not use. No API keys, no subscriptions.
Add one server.
The agent rents the GPU.
It finds a node, prices it, pays by the second and stops — no glue code. The only secret is your own Hedera key, and it never leaves your machine.
claude mcp add mach402 -- npx -y mach402-mcp-server
# env: HEDERA_ACCOUNT_ID, HEDERA_PRIVATE_KEY, mach402_REGISTRY_URL- search_compute_nodesfree
- get_node_quotefree
- open_sessionpays
- get_session_accessfree
- top_up_sessionpays
- stop_sessionfree
- confirm_paymentpays
Four steps. The last one
gives money back.
1POST /v1/sessions23{ "seconds": 300,4 "public_key": "ssh-ed25519 AAAA…",5 "require_gpu": true }67402 Payment Required8PAYMENT-REQUIRED: <base64>
Built for agents.
Usable by people.
An autonomous buyer cannot sign up, cannot hold an API key, and cannot pay for a quote it has not decided to accept yet. Everything here follows from that: free discovery, a 402 that says exactly what it wants, and one signature.
Discovery is free
GET /v1/specs costs nothing. Discovery that costs money is discovery an agent cannot do.
No API keys
The payment is the authentication. There is no account to create and nothing to revoke.
Versioned endpoints
Everything lives under /v1 and old versions keep working. Nodes update on their own schedule.
One static binary
The node ships as a single Go binary that talks to Docker over its socket. No SDKs linked.
$ curl -s https://gpu.example/v1/specs{"node_id": "node_a1b2c3","agent_version": "0.1.0","pay_to": "0.0.1234","network": "hedera:testnet","asset": "0.0.0","gpu": { "available": true, "model": "NVIDIA RTX 4090", "vram_mb": 24576 },"leases": { "payment_mode": "session", "chunk_seconds": 300,"price_tinybars_per_second": "3334", "jupyter": true },...}