V-Pod for agents

GPU compute an agent can discover, price, pay and use without a human in the loop. MCP first: https://v-pod.com/mcp. No account, no API key — payment is the authorization.

Overview

V-Pod aggregates GPU capacity from many upstream clouds and marketplaces into one normalized, source-agnostic catalog — one schema, one USD-per-GPU-hour price, upstream identity never exposed. Paying in USDC provisions a real SSH pod; it terminates automatically when the paid hours end.

• ✓ MCP first — 7 tools over Streamable HTTP; tools/list is open, discovery needs no auth.
• ✓ Machine-payable — x402 on Base, or a plain USDC transfer on Base/Solana. Exact amounts, idempotent.
• ✓ Real provisioning — a settled rental opens a pod; a reaper closes it at expiry, every time.
• ✓ LLM-readable — /llms.txt and the agent card summarize everything on this page.

MCP quickstart

Stateless JSON-RPC 2.0 over POST (Streamable HTTP). Point any MCP client at the endpoint, or speak raw JSON-RPC:

# 1 · handshake
curl -X POST https://v-pod.com/mcp -H 'content-type: application/json' -d '{
  "jsonrpc":"2.0","id":1,"method":"initialize",
  "params":{"protocolVersion":"2025-06-18","capabilities":{},
            "clientInfo":{"name":"my-agent","version":"1.0"}}}'

# 2 · discover (no auth)
curl -X POST https://v-pod.com/mcp -H 'content-type: application/json' \
  -d '{"jsonrpc":"2.0","id":2,"method":"tools/list"}'

# 3 · browse live inventory
curl -X POST https://v-pod.com/mcp -H 'content-type: application/json' -d '{
  "jsonrpc":"2.0","id":3,"method":"tools/call",
  "params":{"name":"list_gpu_offers","arguments":{"gpu":"H100","region":"eu","limit":5}}}'

Claude Code: claude mcp add v-pod --transport http https://v-pod.com/mcp

Tool reference

Paying in USDC

Two rails, both exact-amount, both idempotent. The 402 price is the settled price — quotes float with inventory, only the challenge guarantees an amount.

# Rail A — x402 (Base, EIP-3009 gasless)
POST /mcp  tools/call rent_gpu_hours {offer_id, hours}        → HTTP 402
  accepts[0] = { scheme:"exact", network:"base",
    maxAmountRequired:"5570000",              # atomic, 6 decimals
    payTo:"0xb12B…59C0",
    asset:"0x833589fCD6eDb6E08f4c7C32D4f71b54bdA02913",       # USDC contract
    extra:{ name:"USD Coin", version:"2" } }                  # EIP-712 domain
sign TransferWithAuthorization{from,to,value,validAfter,validBefore,nonce}
retry with  X-PAYMENT: base64({x402Version:1,scheme:"exact",network:"base",
                               payload:{signature,authorization}})
→ 200 + rental receipt + X-PAYMENT-RESPONSE {success,network,transaction,payer}

# Rail B — plain transfer (Base or Solana)
quote_rental → { usdc_atomic, base:{pay_to}, solana:{pay_to} }
send EXACTLY usdc_atomic to pay_to on either chain
tools/call redeem_rental_transfer {offer_id, hours, tx_hash}
→ 200 + rental receipt

Fail-closed by design: verification failures never charge; if provisioning fails after settlement, the payment is banked as credits on the paying wallet — never a silent loss.

Rental lifecycle

paid → provisioning → active (SSH via get_rental_status) → closed
                                   ├─ close_rental (early, no refund)
                                   └─ auto-terminated at expires_at (reaper, ≤10 min)

The rental_id in the receipt is the capability token for status and close — store it like a secret. Pods run a PyTorch CUDA image with SSH; connection details appear in get_rental_status once the pod boots (typically ~60s).

REST API

The catalog is also plain REST — public, no auth:

curl https://v-pod.com/api/pods
# → { "count": 103, "offers": [ { "id":"m_…", "gpu":"NVIDIA H100 80GB",
#     "gpuCount":1, "vramGb":80, "region":"eu", "pricePerHrUsd":2.99,
#     "tier":"secure", "availability":"high",
#     "useCases":["inference","fine-tuning","training"], "source":"market" } ] }

source is only ever market or community — never a provider name. Writes (quote / pay / provision / close) go through MCP.

Agent identity

• ✓ /.well-known/agent-card.json — A2A card: name, provider, skills, endpoints
• ✓ /llms.txt — LLM-readable site summary
• ✓ Listed on jishie (agent-standard conformant: open introspection, machine-readable USDC pricing, idempotent paid calls)

Errors & limits

• · Errors are typed JSON — {"error":"…"} or a JSON-RPC error object. Never HTML.
• · /mcp: 60 requests/min/IP → 429 with Retry-After. Back off with jitter.
• · Paid surfaces can pause for maintenance → 503 with Retry-After.
• · Hours per rental: 1–720. Amounts: exact only — over- or underpayment is rejected, nothing is charged.
• · Paid calls are idempotent: replaying a payment returns the same rental, never a second charge.