This has been in the works for a long time, and I really should have done it sooner. But regardless, LumaBrowser is now officially open source.
Since its initial launch a year ago, it has racked up thousands of installs and an overwhelming amount of user feedback. It has come a long way since then.
I don’t throw around the term “Anthropic/ChatGPT in a box” lightly—it essentially packages an entire ecosystem into a single project. Trust me, I can already sense the collective eye-roll at hearing that phrase. But it truly comes with almost everything built in, right down to an automated onboarding system that sets up your LLMs, image, music, and voice models for you.
It can even automatically import models straight from existing LM Studio installations (and other tools).
The system automatically loads and unloads models based on your hardware specs and performance, behaving… exactly how you’d expect a chat agent to behave. If you ask it to generate an image, it constructs the prompt, invokes the appropriate tool, and can unload the LLM to load the image model, generate the image, and then swap the LLM right back in.
It can spin up a web server for your local network so anyone on your LAN can access it. It even supports quickly clustering multiple machines together or configuring a domain name if you want to port-forward and share it with friends over the internet. You can also auto-create game modes with shareable links for friends, complete with hooks that call back into the LLM itself.
You can also run one instance of LumaBrowser on a server and another on a client, allowing you to remotely mount its GPUs or directly invoke its running LLMs.
It supports agent creation, scheduled tasks, cron jobs, and event-triggered tasks.
It features an incredibly capable luma CLI agent for coding and desktop tasks… complete with built-in plugins for JetBrains IDEs and VS Code. VS Code is practically embedded right inside it. Or if you want to keep things simple, you can just use its built-in “code mode.”
Not to mention a full suite of roleplay extensions that automatically update scenes, character info, and more.
On top of that… there are features beyond what you’d expect: RAM pinning (keeping models hot in memory between switches if you have extra RAM); a custom llama.cpp build that keeps the chat cache warm (optional, but makes RAM pinning even faster); a custom-trained JEV-like model for routing tool calls…
It supports building dashboards out of live artifacts generated in chat, shipping by default with a custom “hub” project that lets you import and merge calendars from multiple sources like Google and Microsoft 365. You can do the same for task lists like ClickUp (field mappings included, of course), and even scrape/intercept incoming notifications if you want to aggregate them…
Why do all this? Because it brings everything under one roof, giving your local AI assistant full context over your schedules, tasks, and notifications.
There’s also live tab sharing: if you spin up the web server, you can right-click any browser tab to share it… give someone the URL, and they can stream-view that tab with interactive controls enabled… like when you’re ordering Chili’s and want to shoot the tab over to your wife so she can add what she wants… (a modest little plugin, but super handy).
The bottom line is… this is my magnum opus, and honestly, it’s practically overflowing with features.
I hope you all here enjoy it as much as I loved building it. God knows how much blood, sweat, and tears went into this.
https://github.com/amurgola/LumaBrowser
Top Comments
Guilty_Rooster_6708 (⬆51): After seeing so much AI slop lately, reading something this long written by an actual human (I assume) feels weirdly refreshing. I’m currently running a 27B EXL3 + Unsloth Desktop setup as my chat UI, but it lacks an Artifacts feature like Claude’s, so I’m definitely intrigued to give this a spin. Thanks for sharing.
valdev (OP) (⬆25): Haha, glad my human word slop brought some joy. Honestly, I was pretty self-conscious about rambling on like this—conciseness is definitely not my strong suit.
Sirramza (⬆15): I can’t wait for AI to start using the term “human slop.” “Ugh, another app that took 6 months to write. Human slop.”
Useful_Disaster_7606 (⬆11): Wait… it has built-in Playwright-like mechanics? Edit: I’ve been looking for a “browser-first” testing/automation framework like this for ages! Thanks! Definitely checking it out when I have some time.
valdev (OP) (⬆6): Yep! I’ve built in extensions so you can hook into it with Playwright, and you can even use llmFallbacks. For example, if you target #usernameButton, you can add a fallback description like “username button.” If that element can’t be found (say, because the HTML changed), it’ll call an LLM to locate it! Best of all, execution won’t break, and it returns the corrected element.
Exciting_Variation56 (⬆7): Testing this immediately. Have you had any luck running MLX on 24GB of RAM?
valdev (OP) (⬆2): I’ve only tested it on a 48GB Mac, where it ran quite well! To be candid though, I’m lucky enough to have a 5090 and three 3090s… so I mostly run things on that setup. If you run into any issues or have suggestions, please let me know!
valdev (OP) (⬆3): If the artifacts are web URLs, it fully supports them. It runs a full Electron browser under the hood, so as long as it’s pointed to the right path, it can handle it.