Gateway
One inference API for every certified model, with a prepaid balance, budgets, routing rules, and a receipt for every request.
What it is
Inference Bureau sits between your application and the model providers. You call one endpoint, https://api.inferencebureau.com/v1, with the OpenAI SDK you already use. Model slugs match OpenRouter, so a slug like openai/gpt-4o-mini works as written. Models certified for the Anthropic Messages API (every Claude model, and others the gateway converts for) also answer at POST /v1/messages, so the Anthropic SDK works with a base URL change. Models certified for the OpenAI Responses API answer at POST /v1/responses, for responses.create and Codex. POST /v1/embeddings turns text into vectors in the OpenAI shape, and for classification and scoring, POST /v1/evaluate runs TypeSafe Jev.
Every request is checked against your team's Directives, rate limits, and balance before it reaches a provider. Afterwards you can look it up by id to see which provider served it, how many tokens it used, and what it cost.
How a request flows
- Authenticate. The key identifies the team, project, and environment.
- Admit. Directives decide which models and providers are allowed, rate limits are checked, and the worst-case cost is reserved from the team balance.
- Dispatch. A certified provider for the requested model is chosen. If it fails before answering, Dispatch may try another provider for the same model (see Routing).
- Settle. Actual usage replaces the reservation and the unused amount is released. The receipt in Logs shows the final charge.
Concepts
| Term | Meaning |
|---|---|
| Team | Owns the balance, members, and team-wide Directives. |
| Project | Groups environments and keys for one application. |
| Environment | A stage within a project: production or development. Every key belongs to one. |
| Directives | Policy at team, project, environment, or key level: allowed models, limits, and budgets. |
| Dispatch | Routing: which providers may serve a model, and in what order. |
| Logs | Per-request metadata, routing decisions, and costs. Prompts and responses are kept only when you turn on Log content. |
Next steps
- Quickstart: make your first request in a few minutes.
- Migrate from OpenRouter: change two lines.
- Directives: limit models, rates and spend before you hand out keys.
- AI SDK: use Inference Bureau from the Vercel AI SDK.
- Coding agents: point Claude Code, Codex or Cursor at your team's Directives and budgets.
- End users and tags: see and cap cost per end user and per feature.
- Mixed-model sessions: see what running easy turns on cheaper models would save, free and with no code changes.
- Auditor: set up pull request reviews on GitHub.
- List models: the certified models your key can call.
- API reference: every endpoint and parameter.