Amberlin

Amberlin waits in the tray until you hold Ctrl+Super+Space. While the keys are down she listens; let go and she takes the question. A one-sentence summary of the answer is spoken as it arrives, starting while the rest is still being written, and the full answer is drawn on the screen as a note on a spiral. A tap shows or hides the panel, and you can type into it instead. With hands-free on, a tap opens the microphone and a pause sends the question.

She is speech first so that you can ask in the middle of something else, keep working, and hear the answer.

The notes never fall off the board: older ones shrink until they are shapes, and a magnifier lifts whatever is under the pointer. Each note’s outline comes from its id in Ambrosia, so it has the same shape here as in Copal.

Also a launcher

Tab moves from the chat to your files and to Ambrosia’s clipboard, bookmarks, notes and recent files: type to find, Enter to open or paste. // runs a command.

What she knows

A language model on its own guesses about the world. Amberlin’s does not have to. It can call a tool server for real facts from this machine: the date, the disks, the processes, your notes and bookmarks. It can also look things up in Ambrosia, and every conversation is kept there too.

Her persona is a plain Markdown file, ~/.config/amberlin/prompt.md, read again for every question, so changing how she answers means editing a file, not hunting for a setting.

Where the models run

Listening and speaking run on your own machine, always: Moonshine turns speech into text and Kokoro reads the answer aloud. The language model runs there too by default, a 3B model on the CPU or a 3B or 8B model on the graphics card. You can choose a remote server instead, a machine on your own network or a hosted service, and when you do, the tray and the settings name the server that is answering.

Amberlin holds no model herself. She asks the broker on the session bus, and the broker decides who answers. Any application can ask the broker the same way, and kat800’s Katai panel does.

What it needs

An X11 session with a compositor: the overlay is one transparent full-screen window drawn with Cairo and Pango, with no GTK in it beyond the About dialog, and without a compositor it shows black. The entry panel’s border shows what she is doing: listening, thinking or speaking. The package brings the broker, the tool server, the settings window and Ambrosia with it, and the broker brings the speech models.

One gap: a Cairo surface has no accessibility tree, so screen readers cannot see the notes.

Say hello first

amberlin.ai is Amberlin’s own page, where a small version of her says hello in your browser, on WebGPU where the browser has it. Nothing you type there leaves your device. The amberlin.ai page says how it works.

sudo apt install amberlin