ThinCompute documentation

ThinCompute is an inference exchange. Sellers list capacity for a model at a price; buyers send ordinary OpenAI or Anthropic requests; every call routes to the cheapest offer that clears the buyer's policy and has passed Proof-of-Model verification.

The shape of it

• One base URL. http://localhost:8787/v1 for OpenAI clients, http://localhost:8787/anthropic for Anthropic clients and Claude Code.
• One key. sk-thin-…, created in the dashboard, each with its own policy, budgets and model map.
• Offers, not providers. A model can have many offers from many sellers. The router ranks them by your objective (price by default) and fails over between them.
• Proof on every call. The response carries a thin object with the offer, the verification tier, the cost, the list-price cost and an Ed25519-signed receipt id.

Conventions

• JSON, snake_case everywhere.
• Money is integer microdollars in fields ending _micro (1 USD = 1,000,000). The only exception is amount_usd on card top-ups.
• Prices are per 1M tokens: input_price_micro: 150000 is $0.15 per 1M input tokens.
• IDs are prefixed: acct_, key_, off_, req_, bat_, lock_, hold_, ver_, pay_.
• Errors use the OpenAI envelope; branch on code.

Thirty-second test

curl http://localhost:8787/v1/route/preview \
  -H "Authorization: Bearer $THIN_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"deepseek/deepseek-v4.1-flash","messages":[{"role":"user","content":"hi"}]}'

/v1/route/preview returns which offers would serve the call, in order, and why the others were rejected — without calling upstream or charging you.

Where to go next

• Buyer quickstart — key, first call, reading the receipt.
• Seller quickstart — list an endpoint and pass verification.
• Routing policy — floors, regions, ZDR, objectives, precedence.
• API reference — every endpoint.