Embedding

createGateway() returns a Gateway object whose handler is a plain (request: Request) => Response | Promise<Response> function — the WinterCG fetch signature every modern runtime accepts.

import { createGateway } from '@frogbotai/gateway'

const gw = createGateway({
  providers: { openai: {} }
})

Node, via Hono

import { serve } from '@hono/node-server'

serve({ fetch: gw.handler, port: 3939 })

Inside an existing Hono app, mount it under a prefix — bare paths keep working too, so app.mount and direct /v1/... calls both resolve.

app.mount('/v1', gw.handler)

Next.js route handler

// app/api/gateway/[...slug]/route.ts
import { createGateway } from '@frogbotai/gateway'

const gw = createGateway({
  providers: { openai: {} }
})

export const POST = gw.handler
export const GET = gw.handler

Bun and Deno

// Bun
Bun.serve({ fetch: gw.handler, port: 3939 })

// Deno
Deno.serve({ port: 3939 }, gw.handler)

Cloudflare Workers

export default { fetch: gw.handler }

Seed hook context from a request

The handler’s second argument seeds the mutable context bag every hook receives — useful for passing an already-authenticated user into beforeOperation without re-deriving it from headers.

app.post('/v1/*', (c) =>
  gw.handler(c.req.raw, { context: { user: c.get('user') } })
)

This argument only exists on the in-process call — wire traffic hitting the gateway directly can never set it.

Option	Type	Default	Description
`context`	`Record<string, unknown>`	`{}`	Initial value of the `context` bag that every hook phase of this request receives.

Selective route mounting

Mount individual routes instead of the full handler, keyed by bare HTTP path.

app.mount('/v1/chat', gw.routes['/chat/completions'].handler)

In-process model calls

Skip HTTP entirely and resolve a model for direct AI SDK calls — useful when you already have a generateText/streamText call site and just want the gateway’s provider resolution and hooks.

const model = gw.chatModel('anthropic/claude-sonnet-4-5')

gw.operation({ operation, ... }) exposes the full lifecycle (model resolution, hooks, usage aggregation) for building custom in-process call sites — see Hooks.

import { generateText } from 'ai'

const op = gw.operation({
  operation: 'chat.completions',
  model: 'anthropic/claude-sonnet-4-5',
  context: { userId: 'user-123' }
})

await op.start()

try {
  await generateText({ model: op.chatModel(), prompt: 'Hello' })
} finally {
  await op.finish()
}

gw.operation accepts:

Option	Type	Default	Description
`operation`	`'chat.completions' \| 'messages' \| 'responses' \| 'embeddings' \| 'images' \| 'speech' \| 'transcriptions' \| 'videos' \| 'rerank' \| 'evaluate'`	Required	The `operation` value every hook receives.
`model`	`string`	Required	Canonical `<provider>/<model>` ID. The text before the first `/` becomes the hooks’ `provider`.
`context`	`Record<string, unknown>`	`{}`	Initial value of the `context` bag shared by every phase.
`request`	`Request`	—	The HTTP request the operation came from, passed to `beforeOperation`. Omit it for calls that don’t start from a request.

The returned operation has one requestId and context for the whole lifecycle:

Member	Type	Description
`requestId`	`string`	The operation’s `req_<uuid>` ID.
`context`	`Record<string, unknown>`	The shared hook context bag.
`start()`	`() => Promise<void>`	Runs `beforeOperation` hooks. A throwing hook rejects the promise; call it before any model call.
`finish(result?)`	`(result?: { finishReason?: string; usage?: HookUsage; error?: unknown }) => Promise<void>`	Runs `afterOperation` hooks once; later calls do nothing. Values you pass replace the finish reason, summed usage, and last error collected from the model calls.
`chatModel()`, `embedModel()`, `imageModel()`, `videoModel()`, `speechModel()`, `transcribeModel()`, `rerankModel()`, `evaluationModel()`	`() => Model`	Resolve `model` as an AI SDK model of that kind. Each call through it runs `beforeUpstream`, then `afterUpstream` or `afterError`, with the operation’s `requestId` and `context`.

gw.operation runs only the hooks in your config. The gateway’s built-in logging, tracing, and metrics hooks run on HTTP requests, not on in-process operations.

Start here

Basics

Managing Data

Features

Ecosystem

Deployment

Node, via Hono

Next.js route handler

Bun and Deno

Cloudflare Workers

Seed hook context from a request

Selective route mounting

In-process model calls

​Node, via Hono

​Next.js route handler

​Bun and Deno

​Cloudflare Workers

​Seed hook context from a request

​Selective route mounting

​In-process model calls

Node, via Hono

Next.js route handler

Bun and Deno

Cloudflare Workers

Seed hook context from a request

Selective route mounting

In-process model calls