6.0 KiB
LLeMbas Desktop
The LLeMbas agent on this computer, and the chats on your LLeMbas server, in an app of its own.
- The agent, built in, by default. LLeMbas Desktop carries the LLeMbas CLI and runs it in the background as this computer's device. Agent chats work in your project folders here, without a terminal -- start them, follow them and answer them in the app.
- Your server's chats beside them: the same account, models and permissions as in the browser.
- Two explorers, kept apart: the agent's working directory on this computer, and your files on the server.
- A viewer for what you open and for what the agent shows you: Markdown drawn, code highlighted, text, and HTML in a sealed frame that can reach neither the app nor the network.
- Notifications when a reply or an agent's work finishes while you are elsewhere.
- Its own window and tray: close the window and the agent keeps running in the tray.
There are no admin pages: an administrator uses the server's web interface for that.
Beta (1.0). Needs a LLeMbas 2.0 server or later; an older one is refused with a message saying so. Linux first (
.deband AppImage, x86-64). Windows and macOS later.
Install
Download the .deb or the AppImage from the
latest release. Check it against
SHA256SUMS in the same release.
sudo apt install ./llembas-desktop_*_amd64.deb # Debian, Ubuntu
chmod +x llembas-desktop_*.AppImage && ./llembas-desktop_*.AppImage
Installing a .deb of the version you already have changes nothing (apt calls it "already the
newest version"); sudo apt reinstall ./llembas-desktop_*_amd64.deb installs it again. If the app
does not start, the reason is in ~/.local/share/eu.llembas.desktop/logs/desktop.log.
On GNOME the tray needs the AppIndicator extension; without it, the title bar's menu has everything, and starting the app again brings a hidden window back.
Updates
LLeMbas Desktop looks for a new release when it starts and every six hours. When there is one, an
update button appears in the title bar (and Install LLeMbas x.y.z… in its menu; Check for
updates… otherwise). Installing asks for your password to update the .deb (an AppImage updates
itself), then the app starts again. Every package is checked against the release signing key
built into the app before anything is installed. Betas are not offered as updates.
Use
- Start LLeMbas and enter your server's address. Sign in: the approval page opens in your browser, where you are signed in already; approve the code shown in the app. This sign-in lets the app read and write your chats, as the web interface does.
- Under Settings, Link this computer: a second approval, which makes this computer a device the agent runs on. The agent starts.
- Choose the folders the agent may work in (Settings, Add folder…); a new computer allows none until you do.
- New agent chat: pick one of those folders, a mode and a model, and write.
The agent is the LLeMbas CLI's lembas service, using the CLI's own configuration
(~/.config/lembas). The app's copy of the CLI lives in its own directory
(/usr/lib/llembas-desktop/), not on your PATH. If you also installed the CLI and run
lembas service yourself, the app sees it and does not start a second one: they are the same
device.
What can reach what
- The app's page is its own (
ui/), and it is the only content with access to the app. No page of the server is ever loaded into the window, and the window goes nowhere else. - The app's sign-in token stays in the app's Rust half (
account.jsonin the app's config directory, readable by you only); the page asks for the server's/api/v2through it and never holds the token. - The explorer reads only inside the agent's folders.
- HTML is shown in a sealed frame on an address of its own: scripts in it run, but it has no access to the app, no network and no way to navigate anywhere.
- Links in a reply open in your browser, and only web and mail links.
Build
Needs Rust (stable) and, on Linux, WebKitGTK 4.1 and an AppIndicator library:
sudo apt install libwebkit2gtk-4.1-dev libayatana-appindicator3-dev librsvg2-dev libxdo-dev
scripts/fetch-cli.sh # the bundled CLI, checked against its release signature
scripts/fetch-tokens.sh # verifies ui/tokens.css against tokens.lock.json
scripts/fetch-vendor.sh # verifies ui/vendor/ against vendor.lock.json
cargo install tauri-cli --version "^2" --locked
cargo tauri dev # or: cargo tauri build
There is no Node and no bundler: the app's page is plain HTML, CSS and JavaScript modules in
ui/. The version of the CLI it carries is in lembas-cli.version.
Three things in ui/ are copied from the LLeMbas server and committed, each pinned:
ui/tokens.css, the server's design tokens:scripts/fetch-tokens.sh <server ref>from a checkout of the server beside this repository (../LLeMbas, orLEMBAS_SERVER), pinned intokens.lock.json.ui/vendor/, the server's vendored CodeMirror build:scripts/fetch-vendor.sh <server ref>checks every module against the server's own pin before copying it, and records both hashes invendor.lock.json. The one change: the modules import each other by file name, since the page has no import map.ui/i18n.js, the page's words in the server's languages:scripts/gen-i18n.py, from the server's catalogues and the script's own table. Run it when a sentence on the page changes.
scripts/check-ui.sh parses the page's scripts and refuses any that sets HTML from a string;
scripts/hygiene.sh keeps private names out of this public repository. CI runs both, and starts
the built app under X11 and under Wayland (scripts/ci-start.sh).
Licence
MIT; see LICENSE. The bundled LLeMbas CLI is MIT too, with its own third-party notices.
CodeMirror and Lezer (in ui/vendor/) are MIT.