bf16.ai Protocol & Gateway Reference
bf16.ai Developer API Documentation
Point any OpenAI SDK, Anthropic SDK, or coding agent at http://localhost:3100 using a sk-bf16-v1-... Bearer key.
1. Supported Wire Endpoints
| Method & Path | Auth Header | Protocol Contract |
|---|---|---|
| POST http://localhost:3100/v1/chat/completions | Authorization: Bearer sk-bf16-v1-... | OpenAI Chat Completions (SSE streaming, reasoning_content, tools, JSON schema). |
| POST http://localhost:3100/v1/responses | Authorization: Bearer sk-bf16-v1-... | Stateless OpenAI Responses API (store: false enforced for Zero-Prompt Retention). |
| POST http://localhost:3100/v1/messages | x-api-key or Bearer sk-bf16-v1-... | Native Anthropic Messages wire format with HMAC-signed thinking blocks and tool_use translation. |
| GET http://localhost:3100/v1/models | Optional / Bearer | Live catalog metadata, 1M context limits, and per-token USD rates. |
2. Optional Control & Telemetry Headers
X-BF16-Max-Price-Per-M-Output: 2.00— Rejects routing to any endpoint whose output rate exceeds your USD ceiling.X-BF16-Include-Cost: true— Appends settled USD cost in response headers (X-BF16-Cost-USD) and final SSE usage frame.Idempotency-Key: <uuid>— Prevents duplicate non-streaming charges across network retries.
3. Zero-Prompt Retention Guarantee
bf16.ai streams tokens directly through memory without persisting request bodies, system prompts, or model completions to disk or databases. Only token counts, HTTP status codes, and nano-USD ledger postings are logged.