Two ways to run it
Local keeps everything in your browser, with your own provider keys. Server signs users in, stores their data per account, and keeps the API keys where a browser can never read them.
Llooma
A (less) minimal LLM chat app. Run it in your browser alone, or host it for your team.
Bring your own models
Ollama, OpenAI, Claude, Infomaniak and anything OpenAI-compatible. Several at once, each with its own label and colour, so you always know which endpoint is about to be billed.
Documents, read in the tab
A PDF, a spreadsheet, a Word file. Parsed in your browser and never uploaded, in server mode too. Optional OCR for scans, and a vision fallback when a page is a picture of words.
Conversations that keep fitting
/compact replaces a long history with a structured summary so it still fits in the window. Nothing is deleted, and one click puts it all back.
Six themes, each with a light and a dark ramp
Section titled “Six themes, each with a light and a dark ramp”Classic, Dracula, Catppuccin, Gruvbox, Nord and Solarized. Pick a palette without committing to a mode, or follow the system without committing to a palette.
Built on
Section titled “Built on”Llooma is a small amount of glue around a lot of other people’s work.
- HollamaMIT
The project Llooma is a fork of, by fmaclen
- officeparserMIT
Reads PDF, Word, PowerPoint and spreadsheets, in your browser
- pdf.jsApache-2.0
Renders the pages of a scanned PDF so a vision model can read them
- Svelte and SvelteKitMIT
The app itself
- Tailwind CSSMIT
Every style in the interface
- Bits UIMIT
Menus, dialogs and popovers that behave on a keyboard
- LucideISC
The icons
- KaTeXMIT
Maths, rendered in a reply
- OllamaMIT
Models running on your own machine
- degoogAGPL-3.0
The default web search backend, self-hosted like everything else here
- StarlightMIT
This documentation site
Not sponsored
When local isn’t an option
Section titled “When local isn’t an option”Running Ollama on your own machine means no third party at all, and that’s what I’d recommend first. It needs a decent GPU though, and no phone has one. If that’s not realistic, someone else runs the model, and it’s worth caring who.
Infomaniak is the one I like: Swiss datacentres they own and run, open-weight models, prompts they don’t log or train on, and the GPU heat goes into warming homes nearby. Any other OpenAI-compatible provider works just as well here.