Skip to main content
The Tulip v1 Gateway is an offer-aware proxy for verified Hugging Face Inference Endpoints. It exposes an OpenAI-compatible chat route while adding model-credit quotes, permit validation, provider routing, metering, and signed receipts.
Hugging Face is the first Gateway adapter, not Tulip’s permanent vendor boundary. Later versions may add other model registries, endpoint vendors, GPU networks, and execution environments.

Base URL

Use the production base URL issued with your Tulip API credentials. The examples refer to it as TULIP_GATEWAY_URL. Send credentials only to that origin.

Capabilities

  • Lists canonical models and public provider offer metadata.
  • Selects the cheapest eligible offer for a request.
  • Reserves a maximum credit charge against an inference permit.
  • Wakes sleeping Hugging Face endpoints when required.
  • Proxies non-streaming and streaming chat completions.
  • Requires exact upstream token usage.
  • Signs offer-bound EIP-712 usage receipts.
  • Onboards provider endpoints without exposing their credentials publicly.
  • Submits completed receipts to the inference escrow in batches.

Compatibility

POST /v1/chat/completions accepts the common OpenAI chat completion fields used by the selected endpoint, including messages, generation limits, tools, response format, and streaming options. Tulip adds:
  • Optional offer in the request body for explicit provider selection.
  • X-Tulip-Permit for credit authorization.
  • Optional X-Tulip-Quote for an existing request-bound quote.
  • Tulip response headers identifying routing and charges.

Public and private data

Public model and offer responses include repository identity, on-chain addresses, operator, payout, status, capabilities, hardware metadata, and verification times. They never include the provider’s endpoint URL or Hugging Face access token.

Error format

Gateway errors use a consistent object:

Authentication and headers

Configure API keys, inference permits, quotes, and provider signatures.