Skip to main content
Your app sends one request to one endpoint with one key. Nito resolves the model and privacy level, routes to an eligible backend, reserves usage, calls the provider, and returns or streams the result.

​
The Request Lifecycle

A simplified flow.

Authenticate

The key in Authorization: Bearer resolves to your organization. Missing, malformed, revoked, or expired keys get 401. Over a rate limit you get 429 with Retry-After and X-RateLimit-* headers.
2

Resolve

The model field selects the model; the catalog entry carries its privacy level. A :confidential suffix asks for the attested variant. A model that cannot serve what you asked for is rejected, never downgraded.

Route

One logical model may have several physical routes. Nito filters by privacy level and the features you sent, then picks a healthy one.

Reserve

Expected usage is reserved against your organization before the provider call. Nothing is charged yet.

Run

Nito calls the provider at the selected level and returns one response or a token stream.
6

Settle

On success the reservation settles for actual usage. A request that fails before producing usage releases it.

​
Public Model Catalog

Model discovery is public. Inference requires an API key and account eligibility. Each entry carries the id you send back, its type, capability flags (supports_reasoning, supports_image_inputs, supports_web_search, and more), and its privacy_tier. The catalog changes over time; treat GET /v1/models as the live source and select by capability at runtime. IDs are vendor/model, for example openai/gpt-5.5 or moonshotai/kimi-k3. See Model ID Convention.

​
Level by Model ID

Available levels appear as separate model IDs and privacy_tier values in the catalog. Nito stores nothing you send or receive at any level. See Privacy Levels. The only suffixes are :confidential and :encrypted. There is no privacy field in the request body, and a level never downgrades: a model that cannot serve it returns an error before any provider call.

​
What One Key Covers

  • Chat at /v1/chat/completions (OpenAI shape) and /v1/messages (Anthropic shape).
  • Web search and web fetch, run by Nito before the model call.
  • File input: PDFs and office documents parsed into text.
  • Reasoning, with effort and summary controls.
Each has a page under Features.

​
Two Rules

  • No fallback. A :confidential call that cannot be served as asked fails.
  • Routes can vary. The same logical model may run on different eligible backends across calls. For Confidential calls, the attestation evidence names what ran.

​
Next

Quickstart

A first request.

Authentication

Keys and headers.

Privacy Overview

The four levels and attestation.