Duet

Collaborative AI sessions - multiple participants, one shared context, real-time streaming.

AI conversations today are single-player. Duet puts everyone in the same AI conversation. Messages from all participants feed into a single context, and the AI streams its response to everyone simultaneously.

How it works

1. Create a session and share the invite link
2. Participants join over WebSocket
3. All messages feed into a single AI context
4. The AI streams its response to everyone in real time

The AI sees everything every participant contributes and addresses people by name.

Quick start

Prerequisites

• Node.js 24+
• pnpm 9+

Setup

git clone https://github.com/joelfickson/joinduet.git
cd joinduet
pnpm install

Configuration

Set your LLM provider API key. At least one is required:

# Anthropic (default)
export ANTHROPIC_API_KEY=sk-ant-...

# Or Google Gemini
export GEMINI_API_KEY=AIza...

# Or OpenRouter (has free models)
export OPENROUTER_API_KEY=sk-or-...

Or use BYOK - provide your API key when creating a session in the UI.

Run

# Start both server and client
pnpm dev

# Or separately
pnpm --filter @duet/server dev   # Server on :8000
pnpm --filter @duet/client dev   # Client on :8001

Open http://localhost:8001, create a session, share the invite link.

Architecture

Participant A                 Participant B
      |                             |
      +---------- WebSocket --------+
                      |
                Session Server
                      |
           +----------+----------+
      Context Store          LLM Pipeline
      (shared, neutral)      (multi-provider)
           +----------+----------+
                      |
            Fan-out via WebSocket

Monorepo structure

packages/
  server/    @duet/server   Fastify + WebSocket, LLM streaming, SQLite
  client/    @duet/client   React Router v7, Zustand, native WebSocket
  shared/    @duet/shared   TypeScript types (events, messages, sessions)

LLM providers

Ships with three providers, selectable per session:

Provider	Default model	Notes
Anthropic	Claude Sonnet 4	Best quality
Gemini	Gemini 2.5 Flash	Fast, good value
OpenRouter	Llama 3.1 8B	Free models available

Environment variables

Variable	Default	Description
ANTHROPIC_API_KEY	-	Anthropic API key
GEMINI_API_KEY	-	Google Gemini API key
OPENROUTER_API_KEY	-	OpenRouter API key
PORT	8000	Server port
DATABASE_PATH	./data/duet.db	SQLite database path
VITE_API_URL	http://localhost:8000	API URL for client
VITE_WS_URL	ws://localhost:8000	WebSocket URL for client
RATE_LIMIT_MAX	30	Max messages per window
RATE_LIMIT_WINDOW_MS	10000	Rate limit window (ms)
RECONNECT_GRACE_MS	60000	Reconnection grace period (ms)

Development

# Install dependencies
pnpm install

# Run dev servers (both)
pnpm dev

# Lint + format (Biome)
pnpm check

# Type checking
pnpm typecheck

# Tests
pnpm test

Tech stack

• Runtime: Node.js + TypeScript (strict mode)
• Server: Fastify, @fastify/websocket, better-sqlite3
• Client: React Router v7 (framework mode), Zustand, TanStack Query
• LLM: Anthropic SDK, Google GenAI SDK, OpenAI SDK (OpenRouter)
• Tooling: Biome (lint + format), Vitest (tests), pnpm workspaces

License

MIT