Routing

POST /v1/chat/completions (and POST /v1/completions) accept OpenRouter-style suffixes on model, plus a virtual auto id. Suffixes are stripped before provider resolve. Caller provider.order and provider.allow_fallbacks are never overwritten.

Suffixes

Suffix	Effect
`:nitro`	Prefer throughput (`provider.sort = "throughput"`) unless you already set `sort`
`:floor`	Prefer price (`provider.sort = "price"`)
`:free`	If the catalog has $0 input+output rows for that model, restrict `provider.only` to those slugs. Otherwise `sort=price` and ignore `openai`, `anthropic`, `azure-foundry`, `google`

Ollama tags such as llama3.2:3b are unchanged — only :nitro, :floor, and :free are stripped.

Auto router

opendoor/auto, openrouter/auto, and auto pick a live serverless or open-weight catalog model, preferring the cheapest healthy option. If the catalog query fails, the gateway falls back to gemma-4-26b-a4b-it, then deepseek-v3.2, then llama3.2:3b.

Restricted API keys are checked against the resolved id: llama-3.1-8b-instruct:nitro matches an allowlist of llama-3.1-8b-instruct. If the key has an allowlist, opendoor/auto picks the cheapest/first live model from that list instead of 403ing because auto is not listed.

You can still pass provider (order, allow_fallbacks, only, ignore). Suffix sort merges in only when sort is unset.

Examples

bash
curl http://localhost:3001/v1/chat/completions \
  -H "Authorization: Bearer $OPENDOOR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"llama-3.1-8b-instruct:nitro","messages":[{"role":"user","content":"hi"}]}'

bash
curl http://localhost:3001/v1/chat/completions \
  -H "Authorization: Bearer $OPENDOOR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"llama-3.1-8b-instruct:floor","messages":[{"role":"user","content":"hi"}]}'

bash
curl http://localhost:3001/v1/chat/completions \
  -H "Authorization: Bearer $OPENDOOR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"opendoor/auto","messages":[{"role":"user","content":"hi"}]}'

Optional app attribution (stored on request metadata when present): HTTP-Referer and X-Title.

Transforms

POST /v1/chat/completions accepts OpenRouter-compatible transforms. The only implemented value is middle-out.

When transforms includes "middle-out" and the prompt is longer than the catalog context_window minus max_tokens, the gateway keeps system messages and the first/last slices of the rest (and clips oversized individual messages). It is opt-in: omitted transforms leaves the prompt unchanged. Models without a catalog context window are not truncated.

bash
curl http://localhost:3001/v1/chat/completions \
  -H "Authorization: Bearer $OPENDOOR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"llama-3.1-8b-instruct","transforms":["middle-out"],"messages":[{"role":"user","content":"hi"}]}'

BYOK

Service tiers