A small, offline chat front-end for your local Ollama models. Double-click to open; it starts the Ollama daemon for you and shuts it down — unloading models from RAM/VRAM — when you close the window.
An independent, unofficial client. Not affiliated with, endorsed by, or sponsored by Ollama. "Ollama" is used here only to say what this talks to.
./install.sh # adds the launcher + icon to your app menuThen launch LocalLM Studio from your application menu, or double-click
run.sh. To remove the launcher again: ./uninstall.sh.
Requirements: ollama on PATH, Python 3.11+, and either GTK3 + WebKit2
(python-gobject, webkit2gtk-4.1) or PySide6. No pip packages, no Electron,
no bundled browser. If neither webview is present the UI opens in your default
browser instead.
- Chat with streaming replies, markdown, and copyable code blocks.
- History in a searchable sidebar; rename, delete, regenerate, edit-and-resend. The sidebar has two tabs: Chats, and Characters for applying a persona to the open chat in one click (+ New character, or ✎ on a row, opens the editor).
- Model picker listing everything
ollama listknows about. - Thinking / effort selector — off, on, or low / medium / high for models
that accept a reasoning effort. Thinking is shown in a collapsible block.
"No thinking" is sent as an explicit
think: false, which genuinely turns reasoning off on models that otherwise default to it. - Parameters per chat: temperature, top-k, top-p, min-p, repeat penalty, context length, max response tokens, seed, stop sequences, keep-alive.
- System prompt per chat, plus reusable characters (personas and scenarios) that are prepended to it.
- 17 themes including Catppuccin, Nord, Tokyo Night, Gruvbox, Dracula, Everforest, Rosé Pine, Solarized, One Dark and Monokai.
The ⚙ panel is split into three tabs — Chat (character, system prompt), Model (a summary of the selected model, sampling, context and limits) and Theme. Everything there applies to the current chat; Save as defaults for new chats pushes it onto every chat you create afterwards.
On launch the app looks for a daemon already listening on OLLAMA_HOST or
127.0.0.1:11434. If it finds one it attaches and leaves it alone. Otherwise it
spawns its own ollama serve and, on quit, evicts every resident model
(keep_alive: 0) before terminating that process.
Its own daemon binds 127.0.0.1:11434 — the address the ollama CLI expects —
so ollama ps, ollama list and the rest work against it with no extra setup.
Set OLLAMA_HOST to choose a different port; only loopback addresses are
honoured, so pointing it at 0.0.0.0 or another machine is refused rather than
obeyed. If the wanted port is held by something that isn't a reachable daemon,
the app falls back to a random free one.
To check whether anything is still running:
pgrep -a ollama # prints nothing once the app has shut it downPrefer that over ollama ps: the CLI reports "could not connect" both when the
daemon is stopped and when it is listening somewhere other than 11434. To find
the live port either way, hover the status dot in the sidebar, or run:
OLLAMA_HOST=127.0.0.1:$(ss -ltnp | awk '/"ollama"/{n=split($4,a,":"); print a[n]; exit}') ollama psQuit from the Quit button, by closing the window, or with Ctrl-C in a terminal — all three run the same teardown. “Unload” frees memory without quitting.
- Nothing is sent anywhere. The only network traffic is loopback, between this app and your local Ollama daemon; the HTTP client is built with an empty proxy handler so a system proxy cannot pull it off-box.
- The UI server binds
127.0.0.1on an ephemeral port. Every/api/call must carry a session token in theX-Auth-Tokenheader — never in the URL. At launch the window receives a single-use bootstrap token that expires after 30 seconds and is redeemed once, in-page, for the session token, so a launch URL recorded anywhere is already spent.HostandOriginare pinned to our own origin, which blocks DNS-rebinding and cross-site requests. - The page is served under a Content-Security-Policy of
default-src 'none'withconnect-src 'self': no remote scripts, styles, fonts or images can load, and the page cannot call out even if something tried. - Model output is rendered by building DOM nodes — never
innerHTML— so markup in a reply cannot become markup on the page. Links are restricted to http/https and open in your system browser rather than inside the app. - The webview runs with an ephemeral data store: no cookies, cache or local storage are written to disk.
- History lives in one SQLite file at
~/.local/share/locallm-studio/chats.db, created0600in a0700directory. Deleting that file deletes everything. - Sampler values from the UI are whitelisted and clamped server-side, and IDs are validated as 32-char hex before touching the database.
app/ Python backend — daemon lifecycle, loopback API, SQLite store, window
web/ The UI: one HTML page, two stylesheets, three scripts
run.sh Launcher
Environment overrides: LOCALLM_STUDIO_BACKEND=gtk|qt|browser forces a webview,
LOCALLM_STUDIO_DEBUG=1 enables devtools and the context menu.
Built on Ollama. The bundled palettes are the work of their respective projects — Catppuccin, Nord, Tokyo Night, Gruvbox, Dracula, Everforest, Rosé Pine, Solarized, One Dark and Monokai — reproduced here as colour values only; see each project for its own terms.
MIT. See LICENSE.