Skip to main content
Spillway is an OpenAI- and Anthropic-compatible AI gateway with a governance brain: hierarchical budgets with hard enforcement, guardrail policies, approval workflows, and cost-aware routing, sitting in front of every model call your organization makes. You point your existing SDK at Spillway instead of the provider directly. Every request is then metered, priced, governed, and logged before — not after — it reaches the provider.

Quickstart

Switch two lines, make your first governed request, and watch a budget block spend before it happens.

Installation

Run Spillway locally: Postgres, migrations, the seed script, and the dev server.

API reference

Every data-plane and control-plane endpoint, generated from the request/response schemas the server actually validates against.

Architecture

The two-plane design: one Fastify process, one Postgres database, hard governance checks before any provider call.

​
The two planes

Local dev serves both planes from http://localhost:3000.
Admin-key authentication isn’t live yet — see the API reference for status. Use a session token (dashboard login) for the control plane until then.

​
What governance actually means here

Most gateways log spend after the fact. Spillway’s guarantees are enforced before the request leaves the building:
  • Budgets that block. A budget in enforce mode returns 402 before the provider is called once a scope’s spend for the period would exceed its limit — no upstream call, no charge, no surprise invoice.
  • Guardrail policies. Match on virtual key, team, model, or metadata; deny, flag, or require approval. Evaluated on every request, not sampled.
  • Approval workflows. A policy can gate a request behind a multi-step, tiered approval chain instead of an outright deny.
  • Cost-aware routing. Routing rules rewrite a model to a cheaper equivalent or a fallback chain, evaluated per request before dispatch.
  • A ledger, not a log. Every request — including every block — is written as a row with its cost, scope, and outcome, so budgets, alerts, and chargeback reports are always reconciled against what actually happened.

​
Where next

1

Read the Quickstart

Switch your SDK’s base URL and key, make one request, then watch a budget block the next one. Five minutes.
2

Run it locally

Follow Installation if you want to run the gateway and dashboard yourself rather than against a hosted instance.
3

Configure governance

Configuration covers budgets, guardrail policies, routing rules, approvals, and teams — the resources you’ll actually create.