A private ChatGPT-style assistant on your own PC. Download, open it, click Easy Setup, and start chatting. No account or API key needed for local models. Free and open source, with every installer published on GitHub Releases.
Image generation, roleplay, voice, and small games on your own graphics card. Download, open it, click Easy Setup, and it installs the image model for you. No Python, no account.
A real browser your agent can drive: REST API on port 3000, a built-in MCP server, and WebDriver and CDP endpoints you switch on in Settings. Download below, or skip to the command-line and Docker options.
Already run LM Studio, Ollama or ComfyUI? Easy Setup adopts the GGUFs you have by hardlink, runs on llama.cpp and stable-diffusion.cpp, and decides whether chat and images share the card or take turns. What you keep and what changes.
Every knob is in the same free download: placement modes, runtimes, quants, tensor split, MoE offload, distributed GPUs. Install below, then open the Setup tab. Advanced page.
| System | File | Size |
|---|---|---|
| Windows 10 / 11 (64-bit) | Installer (.exe)LumaBrowser Setup 1.8.0.exe | |
| Windows portable | Portable (.exe, no install)LumaBrowser-1.8.0-portable.exe | |
| macOS, Apple Silicon (M1 to M4) | App bundle (.zip)LumaBrowser-1.8.1-arm64-mac.zip | |
| macOS, Apple Silicon (disk image) | Disk image (.dmg).dmg disk image | |
| Linux (x86-64) | AppImageLumaBrowser-1.8.0.AppImage | |
| Debian / Ubuntu (x86-64) | Package (.deb).deb package |
Each installer is a normal desktop app. The only speed bumps are the ones every unsigned download hits, and they take one click each. Then the first-run guide takes over.
LumaBrowser.Setup.2.0.0.exe (GitHub swaps the spaces in the file name for dots).%LOCALAPPDATA%\LumaBrowser; you can move that folder later from Settings.Some antivirus products flag new, unsigned installers on sight. If yours quarantines the file, restore it and add an exception for the LumaBrowser folder. The app sends an anonymous version check-in you can turn off in Settings. Read the privacy policy, or the source.
The portable build is a single .exe you can run from any folder, including a USB stick. It keeps its data next to the standard location.
The main macOS download is a .zip of the app, built for Apple Silicon (M1 or later); a .dmg disk image of the same build is in the full list if you prefer one. Intel Macs are not supported from 1.8 onward.
.zip in Downloads. It unpacks to LumaBrowser.app.LumaBrowser.app into your Applications folder.If macOS still refuses, run this in Terminal and try again:
xattr -dr com.apple.quarantine /Applications/LumaBrowser.appVoice input (speech to text) is not available on macOS in this release; everything else, including local chat on Apple Silicon, is.
The AppImage needs a few system libraries that stock Ubuntu 24.04 (and most recent Debian-based distributions) do not ship. Install them first, then make the file executable:
sudo apt install libfuse2t64 libnspr4 libnss3 libasound2t64 libgomp1chmod +x LumaBrowser-2.0.0.AppImage && ./LumaBrowser-2.0.0.AppImageOn Ubuntu 22.04 and older the packages are named libfuse2 and libasound2. On Fedora use fuse-libs nspr nss alsa-lib libgomp. If the app still refuses to start, the error names the missing library; email us the line and we will add it here.
libgomp1 is needed by the chat engine, not the app, so its absence only shows up after Easy Setup has downloaded a model. Install it now and skip that surprise.
Chat uses the CUDA build automatically. Image generation on NVIDIA Linux currently uses the Vulkan build; see the image guide for what to expect.
Everything below is the same free app. These are just faster doors for people who live in a terminal. Full walkthrough on the agent quickstart.
npx lumabrowser startRequires Node 18+. Downloads the portable build for your OS on first run, caches it under ~/.lumabrowser/, and launches it. npm install -g lumabrowser works too.
git clone https://github.com/amurgola/LumaBrowser && cd LumaBrowser && LUMA_VNC_PASSWORD=change-me docker compose up -dBuilds the full app headless from the repository: the desktop in your browser on 6080, the REST API and MCP proxy on 3000. No GPU acceleration in the container. Docker deployment guide.
In Settings, Connect your agent writes the MCP entry into Claude Code, Codex, OpenCode and Cline for you. For any other host, Export MCP Configuration saves a ready mcpServers block (it runs node on the bundled mcp-server.js, so Node.js must be installed). Setup and tool list.
REST API on http://localhost:3000. W3C WebDriver on 9515 and the Chrome DevTools Protocol on 9222 are off by default; switch them on in Settings and Selenium, Playwright, and Puppeteer scripts attach unchanged, seeing only the tabs automation created. API reference.
No feature is held back for a paid plan. The download is the whole product, and the source for all of it is on GitHub.
LumaBrowser is free software under the GNU AGPL v3. There are no paid tiers, no account, and no subscription, and the open edition is the same code and the same installers as everything else we ship. Optional cloud providers charge separately if you choose to connect them.
Using the official builds, at home or inside a company on any number of machines, needs nothing from us. A commercial license exists for organizations that want to modify LumaBrowser or ship private extensions without publishing their changes. It unlocks no features; it is a contract, not a key. How the licensing works.
Yes. It is open source under the AGPL, every feature is in the download above, and there is no trial timer, no account, and no usage limit.
Every installer is a release asset on GitHub, published next to the source code it was built from. The buttons on this page link straight there. The app's built-in updater reads the same releases.
Yes. Clone amurgola/LumaBrowser, run npm install, then npm run dev (Node 20 and a C++ toolchain for the native SQLite module). npm run build packages installers for your platform. Builds you modify must be marked as unofficial; see open source.
The installer is not code-signed yet, so Windows SmartScreen and some antivirus products warn about any new unsigned file. Click More info, then Run anyway. The size and SHA-256 checksum of every file are listed above, matching the ones GitHub shows on the release, so you can confirm you got the file we published.
The app is under 200 MB. The chat model is the big part: about 2 to 5 GB on a laptop, 6 to 20 GB on a gaming PC, 20 to 50 GB on a high-end machine. Easy Setup tells you the exact size before it downloads anything. Add about 6 GB if you turn on images.
Yes. Settings has a models folder setting; point it at another drive and the app moves or re-downloads there. Individual models can be deleted from the Setup tab. Uninstalling the app does not delete models unless you delete that folder.
No. Easy Setup downloads a local model sized to your hardware and everything runs on your own machine. If you would rather use OpenAI or Anthropic models, paste your own key and pay them directly. We never mark up tokens.
An anonymous version check-in (app version, platform, and an installation identifier, so we can count active installs), which you can turn off in Settings, plus update checks and the downloads you ask for. There is no third-party analytics or crash reporting in the app, and the code is open, so you can check. Full details here.
See what's new in 2.0 for release notes, behaviour changes, and early-access features.