LumaBrowser

LumaBrowser replaces the pile of separate tools people run for local AI (a model server, a chat UI, an image generator, a voice stack, a headless browser for agents) with one desktop app. It looks at your GPUs, tells you which models fit before you download them, and runs everything locally. Your own tabs, logins and files stay on your machine, and outside agents like Claude Code, Codex, Cline and OpenCode can drive the browser through MCP.

Get started

1. Install

Download the latest build from GitHub Releases:

Or, with Node 18 or newer, fetch and launch the latest build from the terminal:

npx lumabrowser start

2. Pick a model

Open Models. LumaBrowser reads your hardware and puts a fit badge and a predicted speed on every model before you download it. Search Hugging Face directly, or link models you already have in LM Studio, Ollama or the Hugging Face cache without downloading them again.

3. Give it a task

Open a chat and ask for something that needs the web:

Search the web for the latest llama.cpp release and tell me in a few bullets what changed.

The agent opens its own tabs, reads the pages and answers. Your tabs stay yours.

Which part do I need?

Features

Browser

• Chromium-based browser with tabs, history, bookmarks and a built-in ad blocker (Ghostery engine), plus Chrome extensions: a curated store of ones that work in Electron (Bitwarden, Dark Reader, SponsorBlock, React DevTools and more), any Chrome Web Store link, or an unpacked folder.
• Luma On Demand: a floating voice or text panel on any page that acts on the current tab.
• Dashboard: a grid of live widgets from chats and extensions.

Local LLMs, fitted to your hardware

• llama.cpp runtimes for CPU, CUDA 12, CUDA 13 and Vulkan, MLX on Apple Silicon, and add-on ik_llama.cpp and NInfer runtimes.
• Sizes each model and its KV cache to your video memory, chooses full, partial or CPU offload, and moves MoE experts to the CPU when that is faster.
• Splits one model across several GPUs, or across other machines over llama.cpp RPC.
• Optional local API that speaks both the OpenAI and Anthropic formats on http://127.0.0.1:8317. Off by default.
• Also works with any OpenAI-compatible endpoint (LM Studio, Ollama, vLLM) and with Anthropic.

Image, video, music and voice

• Images: FLUX.1 Schnell, FLUX.2 klein, Qwen-Image and Qwen-Image-Edit, Z-Image Turbo, Chroma, SDXL, SD 1.5, with LoRA support and import from Hugging Face, a URL or a local file.
• Video: Wan 2.2 (text and image to video), LTX-Video, AniSora, MiniMax-H3.
• Music: MiniMax-Music3 (CUDA, Linux or WSL2, about 34 GB of video memory).
• Speech-to-text: Parakeet, Qwen3-ASR and whisper.cpp.
• Text-to-speech: Kokoro, Pocket TTS and Piper voices, plus Chatterbox voice cloning.

Agents and coding

• Chat with tools: browsing, page reading, file and image tools, a knowledge base over your PDFs and documents (BM25 plus optional embeddings).
• Agents: named sub-agents with their own prompt, tools, model and knowledge, callable from chat, the API or MCP.
• Code mode: a coding agent that also powers the luma CLI and the IDE plugins, and can build and hot-install LumaBrowser extensions.
• Tool Forge: the agent writes, tests and publishes new sandboxed tools that only get the network access they declare.
• Connect outside MCP servers over stdio, streamable HTTP or SSE.

Automation and triggers

• Forward web notifications from any tab to a webhook.
• Watch a page, or one element on it, and get a notification or webhook when it changes.
• Capture network responses matching a URL pattern and forward them.
• Recurring AI tasks with an optional response schema and result webhook.
• Start an agent run from an incoming webhook (Slack and GitHub presets), a file or folder change, a notification or a page change, with filters, cooldowns and quiet hours.
• Push to your phone through ntfy.

Sharing

• Share your models, image and voice generation, agents and (opt-in) GPUs with your other devices. PIN pairing, mDNS discovery, pinned TLS.
• Optional web app (PWA) for phones and tablets on your network.
• Share a live tab by link; viewers can watch or take control.

Extensions

• Everything above the core is an extension. Write your own with a manifest.js and a main.js, and add routes, MCP tools and UI. See Documentation/core/shell/extensions.

Connect your AI agent

MCP

LumaBrowser can write its MCP entry into Claude Code, Cline, Codex and OpenCode for you from Settings, or export the config to a file. By hand it looks like this:

{
  "mcpServers": {
    "luma-browser": {
      "command": "node",
      "args": ["<install dir>/resources/app.asar.unpacked/mcp-server.js"],
      "env": { "LUMA_API_PORT": "3000" }
    }
  }
}

The server starts LumaBrowser if it is not running. Tools cover browsing, tabs, page reading and extraction, agents, timed tasks, page monitors, and any MCP servers you have connected to LumaBrowser.

WebDriver and CDP

Both are off by default. Turn them on in Settings, then:

# Selenium (W3C WebDriver), default 127.0.0.1:9515
driver = webdriver.Remote("http://127.0.0.1:9515", options=options)

// Puppeteer or Playwright (Chrome DevTools Protocol), default 127.0.0.1:9222
const browser = await puppeteer.connect({ browserURL: 'http://127.0.0.1:9222' });

Clients only see the tabs automation created, never your own. Either driver can ask the LLM to resolve a selector that stopped matching.

REST and local model APIs

The full API reference (REST, MCP tools and the OpenAI and Anthropic compatible local API) lives at lumabyte.com/apis.

System requirements

Tiers use your largest single card, since cards do not add up for most workloads (LLMs are the exception: they can split across cards).

Privacy

Your browsing, chats, models, generated media and files stay on your machine. Nothing you type into a chat is sent anywhere unless you choose a cloud model.

What does leave the machine, and how to turn it off:

No feature depends on the usage ping.

Security

Browser automation hands an agent your browser. Anything an MCP, WebDriver or CDP client can see, the model behind it can see. Connect only agents you trust, and keep the API, WebDriver and CDP ports on localhost (the default).

The REST API on port 3000 does not require a key by default. Set an API key in Settings before you expose it beyond localhost.

Docker

The Docker image runs the full app headless and serves the desktop in your browser over noVNC. It is meant for servers and agent hosts.

LUMA_VNC_PASSWORD=change-me docker compose up -d

• http://localhost:6080: the app, throug