@inbrowser/relay

@inbrowser/relay is a resumable LLM inference relay. It is a pure transport: it wraps @inbrowser/resumable with Web-standard request handlers, framework adapters, and a reconnecting browser-safe client. The relay does not own any providers — it consumes ModelClient factories from @inbrowser/model and serves them resumably over HTTP.

The relay's primary value is resumability. A backgrounded browser tab, network drop, or stream handler handoff does not have to lose the events already produced by an in-flight generation. The relay constructs a ModelClient for each request, runs its .chat() server-side, writes the streamed ModelEvents to a durable event log, and clients reconnect from their last received offset.

What It Provides

• createRelay, which exposes handleStart(request) and handleStream(request, ctx) as Web Request to Response handlers.
• A provider lookup table of ModelClientFactorys: the relay calls factory({ apiKey, model }) per request to build a ModelClient, then drives its .chat(req, signal) and stores the ModelEvents it streams.
• SSE helpers shared by the relay, any custom ModelClient, and the reconnecting client.
• Astro and Express adapters.
• createResumableClient, which starts a job, tails the SSE stream, and reconnects with from=<offset> when a stream drops.

The model contract (ModelClient, ModelEvent, ModelRequest, ModelMessage, ToolSpec, ModelUsage) lives in @inbrowser/model/contract. The relay re-exports those names for the registration site, and the cloud provider factories (geminiModelClient, openrouterModelClient, requestyModelClient, anthropicModelClient, ollamaModelClient, claudeCliModelClient, claudeCodeModelClient) live in @inbrowser/model/providers/*.

Quick Start

Import the provider factories from @inbrowser/model and register them in the providers map. Each cloud provider factory already matches ModelClientFactory (its config is { apiKey?, model }), so it can be registered directly:

import { createRelay, type ModelEvent } from '@inbrowser/relay';
import {
  geminiModelClient,
  openrouterModelClient,
} from '@inbrowser/model';
import {
  createRtdbJobStore,
  serviceAccountTokenProvider,
} from '@inbrowser/resumable/rtdb';

const relay = createRelay({
  store: createRtdbJobStore<ModelEvent>({
    url: process.env.RTDB_URL!,
    auth: serviceAccountTokenProvider({ keyFile: './sa.json' }),
    rootPath: 'inference_jobs',
    defaultTtlMs: 7 * 24 * 60 * 60 * 1000,
  }),
  providers: {
    gemini: geminiModelClient,
    openrouter: openrouterModelClient,
  },
});

// POST an inference request and allocate a job:
// await relay.handleStart(request)

// Stream the durable event log as SSE:
// await relay.handleStream(request, { jobId, from })

The relay preserves and replays the event log. It does not automatically restart an upstream provider call if the process running that provider is killed.

The relay does not choose URL paths for you. Common route shapes are:

• POST /api/inference/job - call relay.handleStart(request).
• GET /api/inference/job/:id/stream?from=N - call relay.handleStream(request, { jobId: id }).

API Keys

The relay supports two modes, configurable per provider.

BYOK (default). The client sends apiKey in the request body and the relay forwards it to the provider. A missing key is a 400. This is the right mode when the end user brings their own key (the playground case).

Server-managed. List the provider in apiKeys and the relay resolves the key itself, so the browser never carries it on the wire. This is the "your app, your bill" mode.

import {
  geminiModelClient,
  anthropicModelClient,
  ollamaModelClient,
} from '@inbrowser/model';

const relay = createRelay({
  store,
  providers: {
    gemini: geminiModelClient,
    anthropic: anthropicModelClient,
    ollama: ollamaModelClient,
  },
  apiKeys: {
    gemini: () => process.env.GEMINI_API_KEY ?? '',
    anthropic: () => process.env.ANTHROPIC_API_KEY ?? '',
    // ollama omitted, so it stays BYOK (the client supplies its base URL).
  },
});

A client that sends a non-empty apiKey for a server-managed provider gets a 400, so a forgotten BYOK field cannot silently leak to the wire. If a resolver throws, handleStart returns 500 and no job is created.

The function form receives the raw Request, so the key can be derived from an Authorization header the browser already sends (the browser carries its own user token, never the provider key), a session cookie, or a per-user store:

apiKeys: {
  anthropic: async ({ request }) => {
    const userId = await getUserIdFromSession(request);
    const key = await db.getUserKey(userId, 'anthropic');
    if (!key) throw new Error('no anthropic key for user');
    return key;
  },
}

Client

import {
  createResumableClient,
  installBrowserLifecycle,
} from '@inbrowser/relay/client';

const client = createResumableClient({
  startUrl: '/api/inference/job',
  streamUrl: (jobId, from) =>
    `/api/inference/job/${encodeURIComponent(jobId)}/stream?from=${from}`,
  installLifecycle: installBrowserLifecycle(),
});

for await (const event of client.stream({
  provider: 'gemini',
  model: 'gemini-3-flash-preview',
  messages: [{ role: 'user', text: 'Hello' }],
  tools: [],
  apiKey: userApiKey,
})) {
  // Render text, thinking, tool calls, usage, or errors.
}

Framework Adapters

• Hono, Bun, and Cloudflare Workers can call the Web-standard relay handlers directly.
• Astro uses createAstroRoutes(relay) from @inbrowser/relay/adapters/astro.
• Express and Cloud Functions Gen 2 use createExpressHandlers(relay) from @inbrowser/relay/adapters/express.

Documentation

The documentation follows the Diataxis approach: each page serves one kind of user need.

• Tutorial: create a relay with a fake provider - learn the relay flow without real API keys.
• How to wire a web app - connect server routes and the reconnecting client.
• How to write a provider - implement a ModelClient for an upstream LLM and register it without changing the relay.
• API reference - facts about exports, types, handlers, adapters, the provider contract, and SSE.
• How the relay works - the design rationale and reconnection model.

Package Exports

• @inbrowser/relay - relay factory, transport types, the re-exported model contract types, and the ModelClientFactory type for the registration site. Providers are NOT exported here — import the factories from @inbrowser/model.
• @inbrowser/relay/sse - SSE reader and encoder helpers.
• @inbrowser/relay/adapters/astro - Astro route adapter.
• @inbrowser/relay/adapters/express - Express-compatible adapter.
• @inbrowser/relay/client - reconnecting client and browser lifecycle helper.