# AI Gateway

Route model calls through one audited Ferrite egress boundary with provider selection, attribution, budgets, and streaming.

Ferrite AI Gateway gives applications one model-call interface while keeping
provider credentials, outbound authority, budgets, and usage evidence at a
governed boundary.

## Add AI Gateway to a stack

Deploy `apps/ferrite-ai-gateway` as a private sibling and grant the calling app
only the queue or API capability required to submit a model request. Provider
hosts belong in the gateway’s egress allowlist, not every consumer app.

Use `packages/ai-budgets` to attach account, tenant, application, model, and
purpose attribution to each request. Reject a call before provider egress when
the applicable budget cannot admit it.

## Stream model output

The Gateway’s audited egress path supports chunk-by-chunk streaming, so tokens
can reach the app without buffering the complete response. Bound request size,
provider timeout, total response bytes, and retry behavior.

## Protect credentials and prompts

- Store provider credentials as write-only Ferrite secrets.
- Permit only declared provider origins, methods, paths, and headers.
- Do not log bearer tokens or raw sensitive prompts.
- Attribute every charge before returning success.
- Keep retries idempotent and avoid retrying a partially consumed stream unless
  the provider contract supports it.

## Observe usage

Application usage and gateway attribution remain separate evidence. Read app
requests, bytes, latency, and credits in the Dashboard or with the MCP `usage`
tool; inspect gateway model/provider totals through its own bounded read model.

See [Secrets & Egress](/developers/secrets-egress/) and
[Analytics & Usage](/developers/analytics/).