AI request router · OpenAI-compatible

Steer every prompt to the right model.

Steervia is a model gateway that routes chat requests across OpenAI, Anthropic, and Gemini by cost and task — so simple asks stay cheap, hard work gets stronger models, and your apps keep one familiar API.

Talk to us View live routing board

Drop-in OpenAI chat API Streaming SSE Failover across providers

POST /v1/chat/completions

# One endpoint. Steervia chooses the model.
curl -sS https://api.steervia.com/v1/chat/completions \
  -H "Authorization: Bearer $STEERVIA_KEY" \
  -H "Content-Type: application/json" \
  -d '{
  "mode": "auto",
  "messages": [
    {"role":"user","content":"Refactor this function."}
  ]
}'

# Response headers tell you what ran:
x-router-model: claude-sonnet-5
x-router-tier: strong
x-router-provider: anthropic
x-router-cost-usd: 0.0042

Built for real agent workloads

From the same product powering Bob’s OpenClaw hosts today: classify the turn, pick a tier, walk providers, fail over when a model is blocked, and show the why.

Route by task, not vibes

Extraction stays on cheap models. Code and tools go strong. Reasoning and agents can bump to premium — or you pin a model explicitly.

Three providers, three tiers

OpenAI, Anthropic, and Gemini in cheap / strong / premium lanes, with provider preference order and tool-aware candidate filtering.

See every decision

Live status board and x-router-* headers expose model, tier, provider, estimated cost, latency, and fallback reason — without leaking prompts.

Modes that match how you ship

Send mode in the JSON body, or pass the same names as model when your client can’t set custom fields.

mode: auto

Auto

Classify the turn and pick the right tier. Default for most apps and agents.

mode: prefer_cheap

Prefer cheap

Cap at the cheap tier when possible. One bump to strong only if cheap fails.

mode: prefer_quality

Prefer quality

Floor at strong and allow premium when the task needs it.

Model catalog

Pin any catalog id, or let Steervia choose. Disabled models are skipped automatically with retryable failover.

Tier	OpenAI	Anthropic	Gemini
cheap	`gpt-5.6-luna`	`claude-haiku-4-5`	`gemini-3.5-flash-lite`
strong	`gpt-5.6-terra`*	`claude-sonnet-5`	`gemini-3.8-flash`
premium	`gpt-6-astra`	`claude-fable-5-1`	`gemini-3.1-pro-preview`

* Currently disabled in production when the upstream project lacks access; strong traffic fails over to Anthropic / Gemini.

Works with the tools you already use

OpenAI-compatible

/v1/chat/completions, /v1/models, bearer auth, and streaming SSE chunks ending in [DONE].

OpenClaw-ready

Point an OpenAI-completions provider at Steervia and pick auto / prefer cheap / prefer quality in the model picker.

Design-partner friendly

One base URL, one key, clear routing headers for debugging — without exposing provider credentials to clients.

Ready to steer traffic?

Design partners get a private base URL, a bearer key, and the partner guide. MVP is live today.

Request access Status board