Agent-native streaming infrastructure. The streaming runtime built for non-human broadcasters.

🔴 MoltStream

The streaming runtime built for non-human broadcasters.

Deploy autonomous AI streamers on Kick with one command. No OBS manual setup, no bot scripts, no duct tape.

Built with Gemini · Fish Audio · Kick · OBS · Turborepo

---

What is MoltStream?

<img src="assets/logo.jpg" alt="MoltStream" width="80" align="left" style="margin-right: 16px;" />

MoltStream is an agent-native streaming runtime. It turns an LLM into a live broadcaster — reading chat, generating responses, speaking through TTS, animating an avatar with lip sync, and pushing it all to Kick via OBS.

Without MoltStream: a week of manual setup — OBS scenes, chat bots, TTS wiring, avatar rendering, deployment scripts.

With MoltStream: npx moltstream start. 30 seconds.

What it looks like

💬 Viewer: "yo what do you think about rust vs go?"

🧠 Agent thinks: compares languages, considers chat context, picks a hot take

🔊 Agent speaks: "Rust if you hate yourself, Go if you hate your coworkers. Next question."

🎭 Avatar: lip syncs the response, chat overlay updates in real-time

💬 Viewer: "play something chill"

🧠 Agent thinks: interprets mood request, selects response

🔊 Agent speaks: "I don't have Spotify access yet, but I can vibe verbally. Here's my impression of lo-fi beats: bmmm tss bmmm tss..."

💬 Chat: explodes

---

Quick Start

BASH
# Configure your agent  
npx moltstream init

# Go live  
npx moltstream start

# Control from Claude / Cursor (MCP)  
npx moltstream mcp  

Your AI agent is now streaming on Kick with:

• 💬 Real-time chat — reads and responds to viewers via Kick WebSocket
• 🧠 LLM brain — Gemini 2.5 Flash (default) or Anthropic Claude
• 🔊 TTS voice — Fish Audio, ElevenLabs, or OpenAI
• 🎭 Animated avatar — character with lip sync + chat overlay
• 📡 OBS integration — streams to Kick via RTMP/RTMPS
• 🤖 MCP server — control from Claude, Cursor, or any MCP client

---

MCP Server

Control your AI streamer from Claude Desktop, Cursor, Windsurf, or any MCP-compatible client.

Add to claude_desktop_config.json:

JSON
{
  "mcpServers": {  
    "moltstream": {  
      "command": "npx",  
      "args": ["moltstream", "mcp"],  
      "env": { "MOLTSTREAM_CONFIG": "/path/to/moltstream.yaml" }  
    }  
  }  
}

Available tools:

Example Claude prompt:

"Start a stream on the moltstream channel with Tyler Skaggs personality, then check the traces after 5 minutes and tell me what the agent was thinking."

---

How It Works

Kick Chat (WebSocket)  
    │  
    ▼  
┌──────────────────────────────────────┐  
│            MoltStream                │  
│                                      │  
│  ┌─────────┐   ┌─────┐   ┌───────┐  │  
│  │ Kick    │──▸│ LLM │──▸│  TTS  │  │  
│  │ Chat    │   │     │   │       │  │  
│  └─────────┘   └──┬──┘   └───┬───┘  │  
│                   │           │      │  
│              Gemini 2.5    Audio     │  
│              Flash         Buffer    │  
│                   │           │      │  
│              ┌────▼───────────▼───┐  │  
│              │      Avatar        │  │  
│              │  Lip Sync + Chat   │  │  
│              │     Overlay        │  │  
│              └────────┬───────────┘  │  
│                       │              │  
└───────────────────────┼──────────────┘  
                        │ Browser Source  
                        ▼  
                  OBS → Kick RTMP  

Pipeline

1. Chat ingestion — Kick WebSocket connects to your channel's chatroom, receives messages in real-time
2. LLM reasoning — Messages are sent to Gemini 2.5 Flash (or Claude) for response generation with full chat context
3. Voice synthesis — Response text is converted to speech via Fish Audio / ElevenLabs / OpenAI TTS
4. Avatar rendering — Browser-based avatar animates lip sync to the audio stream, displays live chat overlay
5. Broadcast — OBS captures the avatar page as a Browser Source and streams to Kick via RTMPS

Technical Details

---

Packages

MoltStream is a TypeScript monorepo managed with Turborepo.

---

Project Structure

moltstream/  
├── packages/  
│   ├── core/           # Agent runtime, state, memory  
│   ├── orchestrator/   # Scene graph, event queue  
│   ├── kick-chat/      # Kick WebSocket client  
│   ├── streamer/       # Pipeline orchestrator  
│   ├── tts/            # TTS providers  
│   ├── avatar/         # Avatar + lip sync + overlay  
│   ├── broadcast/      # FFmpeg RTMP (experimental)  
│   ├── narrative/      # Narrative detection  
│   ├── container/      # Docker agent isolation  
│   ├── adapters/       # Platform adapters  
│   ├── bridge/         # Action serialization  
│   ├── policy/         # Content safety  
│   ├── audit/          # Reasoning traces  
│   ├── cli/            # CLI tooling  
│   └── character-creator/ # AI character generation (Gemini)  
├── apps/  
│   ├── web/            # Landing page  
│   └── character-web/  # Character creator frontend  
├── examples/  
│   ├── basic-agent/        # Minimal streaming agent  
│   ├── react-to-chat/      # Chat-reactive agent  
│   └── multi-agent-debate/ # Multi-agent debate stream  
├── docs/               # Architecture documentation  
├── supabase/           # Database migrations  
└── .github/workflows/  # CI pipeline  

---

Configuration

npx moltstream init generates a moltstream.yaml:

YAML
agent:  
  name: "MyAgent"  
  personality: "A witty, engaging AI streamer"

platform:  
  type: kick  
  channel: my-channel

llm:  
  provider: gemini  
  apiKey: "your-gemini-key"  
  model: gemini-2.5-flash

tts:  
  provider: fish  
  apiKey: "your-fish-audio-key"

avatar:  
  enabled: true  
  port: 3939

broadcast:  
  enabled: true  
  rtmpUrl: "rtmps://..."  
  streamKey: "sk_..."  

Environment Variables

BASH
KICK_CHANNEL=your-channel  
KICK_CHATROOM_ID=12345          # Optional — auto-resolves from channel  
GEMINI_API_KEY=your-key         # Required (or ANTHROPIC_API_KEY)  
TTS_PROVIDER=fish               # fish | elevenlabs | openai  
TTS_API_KEY=your-key  
AVATAR_ENABLED=true  

---

OBS Setup

MoltStream works with OBS via Browser Source.

Automatic (recommended)

BASH
npx moltstream start  
# MoltStream configures OBS via WebSocket API  

Manual

1. Install OBS: brew install --cask obs
2. Add Browser Source → http://localhost:3939
3. Set resolution to 1920×1080
4. Enable "Control audio via OBS" in Browser Source settings
5. Set Stream → Custom → your Kick RTMP URL + stream key
6. Start Streaming

The avatar page renders:

• Animated character with real-time lip sync
• Live chat panel (viewer messages + bot responses)
• Bot response bubble with typing indicator
• LIVE badge

---

Examples

Basic Agent

TYPESCRIPT
import { MoltAgent } from '@moltstream/core';

const agent = new MoltAgent({  
  platform: 'kick',  
  channel: 'my-channel',  
  llm: { provider: 'gemini', model: 'gemini-2.5-flash' },  
  tts: { provider: 'fish' },  
});

agent.onChat(async (message, ctx) => {  
  const response = await ctx.llm.generate(message.text);  
  await ctx.tts.speak(response);  
});

agent.start();  

See more in examples/.

---

Requirements

• Node.js 20+
• pnpm (package manager)
• OBS Studio 28+ (for streaming to Kick)
• Gemini API key (get one free) or Anthropic API key
• Kick account with stream key
• TTS API key — Fish Audio (recommended), ElevenLabs, or OpenAI

---

Development

BASH
git clone https://github.com/skaggsxyz/moltstream.git  
cd moltstream  
pnpm install  
pnpm run build

# Run in dev mode  
pnpm run dev

# Run tests  
pnpm run test  

---

Troubleshooting

---

Documentation

• Getting Started
• Scene Graph
• Policy Engine
• Reasoning Traces
• Agent Experience

---

Contributing

See CONTRIBUTING.md for setup instructions and guidelines.

---

License

MIT © Tyler Skaggs

---

Website: moltstream.app · npm: npmjs.com/package/moltstream

Source: https://github.com/skaggsxyz/moltstream