Strawberry

A desktop mascot for Linux: a small cel-shaded crab that sits on top of your windows and reacts to what happens on the machine. She comments on notifications, dances to the beat of whatever is playing, cheers your git commits, and answers when you talk or type to her. Every model runs on your own computer through Ollama; nothing is sent anywhere.

What she does

• Notifications. She sees each desktop notification and reacts to it in a line of her own. By default she knows only the app and the sender, never the message (see Privacy).
• Music. Skip, previous, pause, resume, volume and "what song is this" work for any desktop player (Spotify, VLC, Rhythmbox, mpv, a browser tab) over MPRIS, with no account and no setup. She dances to the actual beat of the playing track.
• Git. With strawberry git-hooks install she reacts to your commits and pushes.
• Talking. strawberry hotkey binds Super+Shift+Space on GNOME: press it, speak, and she answers. Or right-click her and choose Chat with Strawberry… (or press T) to type. A plain music command is done in well under a second; anything else goes to the larger local model.
• Tools. You can give her MCP servers (a calendar, notes, Spotify) in the config. None is configured out of the box. See ADAPTERS.md.
• A tray icon. The 🍓 in the top bar shows her state (idle, listening, thinking, talking) and has the same menu as right-clicking her: show/hide, chat, mute, quiet hour, volume, skin, top hat, settings file, restart, quit.

Requirements

• Linux only for now; Windows and macOS are not supported yet.
• X11 or XWayland (GNOME and KDE on Ubuntu, Fedora, Arch and similar, 2022 or later).
• PipeWire (pw-record, pw-dump) and systemd --user.
• A tray that speaks StatusNotifierItem. On GNOME that is the AppIndicator extension, which Ubuntu ships enabled. Without it she still works; her right-click menu has everything.
• Python 3.12 or later, and uv or pipx.
• Ollama. strawberry setup offers to install it.
• A GPU. The default models are sized for a 24 GB NVIDIA card. Smaller cards work with a smaller brain model, and with no usable GPU she runs on the small model only (setup explains the choice for your card). AMD cards work for the models through Ollama's ROCm support; speech recognition then runs on the CPU, because its GPU engine is CUDA only. Intel graphics count as no usable GPU.

Install

uv tool install strawberry-crab        # or: pipx install strawberry-crab
strawberry setup                       # the widget, Ollama and the models, a voice
strawberry                             # she appears
strawberry install                     # optional: start her on login, with the 🍓 tray icon

For speech recognition on an NVIDIA GPU install the CUDA extra instead: uv tool install 'strawberry-crab[gpu]'. On AMD it does not apply.

What each command does:

• strawberry setup can be run again at any time; it keeps the values already in your config. It reads your GPU memory and proposes models that fit, prints each model's licence before pulling it, offers Ollama's installer if Ollama is missing, downloads a Piper voice, downloads the widget binary for your version, backs up and updates config.toml, offers to add the settings your file does not have yet, and offers strawberry install. --yes takes the defaults without asking.
• strawberry install writes one systemd user unit, strawberry-tray.service, plus an autostart entry as a fallback. The tray starts the daemon, the watchers and the widget, and restarts any of them that dies. strawberry uninstall removes both.
• strawberry doctor checks what she depends on and says what is missing and how to fix it: the config, Ollama and each model, GPU memory, whisper and CUDA, the voice, the widget and its version, PipeWire, git, the tray host, whether the notification monitor is allowed, the media players on the bus, the systemd unit (and whether it still points at this install), the git hooks, the daemon, and whether the beat watcher is posting and listening to the player. It only looks: it never starts, stops or restarts anything. It exits non-zero when something is broken. strawberry doctor --talk also runs a short scripted conversation through the running daemon and prints the time each model took. Please include its output in bug reports.

Other commands: strawberry status, stop, restart, say "…", talk, voices, audition, tools, hotkey, git-hooks install|remove. strawberry --help lists them all.

Configuration

strawberry config opens ~/.config/strawberry/config.toml in your editor, creating it with every setting commented if it does not exist. Her right-click menu has the same thing as Settings file…. Run strawberry restart after editing it.

The settings you are most likely to change:

Files she keeps: settings in ~/.config/strawberry/, the widget binary and voices in ~/.local/share/strawberry/, logs and state in ~/.local/state/strawberry/. The XDG variables are respected.

Privacy

Everything runs on your machine. The models run in your local Ollama, speech recognition and her voice run locally, and the daemon listens only on 127.0.0.1. Strawberry makes no network requests of its own except to download what you ask for: the widget binary from this project's GitHub releases, and models and voices from Ollama and Hugging Face during setup. An MCP server you add is its own program and may use the network (the Spotify one talks to Spotify).

What she reads:

• Notifications. A D-Bus monitor sees every desktop notification. By default she uses only the app name and the sender; the message text never leaves the watcher process ([notifications] body = "off"). You can change that, for all apps or per app with body_apps:

  • "off": app and sender only (the default).
  • "react": the small local model reads the message and reacts in its own words, without quoting it.
  • "glance": she first says a plain one-line gist ("Alex asks about lunch at noon."), then her reaction.

  To switch without editing the file, use the tray menu: Message bodies ▸ Off / React / Glance. It writes body into your config (backing it up first) and applies it at once; body_apps entries still win for their apps.

  In every mode a sensitive filter drops one-time codes, sign-in and password-reset messages and bank or card alerts before any model sees them; she says only "Slack sent something private." The filter fails closed: if its model check is unavailable, every body counts as private.

• Media. The title, artist and album the player publishes over MPRIS.

• Audio. Only the playing media player's own output stream, and only to measure the tempo. Nothing is recorded or stored. [beat] enabled = false turns it off.

• Microphone. Only after you press the hotkey (or run strawberry listen), until you stop speaking. The audio is transcribed locally in memory and never written to disk.

• Git. Only if you run strawberry git-hooks install: the repository and branch name, the commit subject, and for a push the remote's name and the number of commits.

No message text is written to any log. The first time she starts she says in her bubble that notification bodies are off and where to change it, and logs the same note.

She remembers the last few exchanges of conversation in memory, for context, and nothing more.

Spotify (optional)

Music control works for every player without Spotify's API. If you want more (play an artist, a playlist, save a track), install the separate Spotify MCP server panuhen/spotify-mcp, authorise it with your Spotify account, and add it to the config as shown in ADAPTERS.md. Strawberry's Spotify adapter then recognises it and adds its reflexes and your artist and playlist names.

Developers

Start with WIRING.md: the message contract, the daemon, the watchers, the gate, the widget and the tray. PACKAGING.md is the plan for the package and the release process; ADAPTERS.md covers MCP servers and adapters; CHANGELOG.md lists what changed in each version.

From a checkout, bin/strawberry stands in for strawberry and runs from the checkout's .venv. Without an exported widget binary it runs the Godot project in widget/ with godot 4.7 from PATH.

uv sync --inexact --group gpu        # never a bare `uv sync`: it prunes the CUDA wheels
bin/strawberry                       # daemon, watchers, widget
.venv/bin/python -m pytest -q        # tests
scripts/build_widget.sh              # export dist/strawberry-widget-<version>-linux-x86_64
scripts/check_phase1.sh              # tests, then a headless widget against a real daemon

Setup and doctor. strawberry setup reads the GPU's VRAM (nvidia-smi; for AMD rocm-smi, else the amdgpu driver's mem_info_vram_total in sysfs) and proposes the models that fit: the tested setup on a 24 GB card (Qwen 27B brain, gemma3:1b, embeddinggemma, whisper medium on CUDA, Piper en_GB-alba-medium), a smaller Qwen3 for 16, 10 and 6 GB, and no brain on the CPU. On an AMD card the brain rows are the same (Ollama runs them on ROCm) and whisper is small on the CPU in every tier, since faster-whisper's GPU engine is CUDA only; Intel and other cards get the CPU tier. Any slot can be typed over. It writes the choices into config.toml (backing it up first and keeping the values already there), prints each model's licence, runs ollama pull for what is missing, downloads the voice, fetches the widget if the installed one is another version, offers to append the settings the file does not have, and offers strawberry install. The models are not part of this package: you download each from its publisher, under its own terms. --yes takes the defaults without asking (--install also installs; --tier 10gb picks a tier). strawberry doctor checks Ollama and each model, the GPU, whisper and CUDA, the voice, the widget version, PipeWire, git, the tray host, the config, whether the bus allows the notification monitor (a BecomeMonitor with a rule that matches nothing, closed at once), the MPRIS players on the bus, strawberry-tray.service (installed, enabled, active, and whether its ExecStart still exists and is this install), the git hooks (core.hooksPath and both hook files, and the strawberry they call), the daemon, and the beat watcher (posting while a player plays, from /health's tempo_age_s, and linked to the player's stream rather than a sink, from pw-dump), prints ✓ / ! / ✗ with a fix per problem and exits 1 if anything is ✗; doctor --talk also times a short scripted conversation through the running daemon, per slot.

Which widget runs. strawberry widget and the tray run the installed binary when it is there, else the checkout's Godot project with godot on PATH, else they say to run strawberry widget --fetch. A binary built from a checkout (scripts/build_widget.sh, needs Godot 4.7.2 and its export templates) can be tried by hand: dist/strawberry-widget-<version>-linux-x86_64 --display-driver x11 -- --ws=ws://127.0.0.1:8770/ws. The widget tells the daemon its version; a source run says dev and is always accepted, a release on another major version is refused (WIRING §1).

Entry points: strawberry (the CLI; strawberry --help), strawberryd (the daemon alone), strawberry-doorway mpris_watch|notify_watch|beat_watch. Files: config and the widget's preferences ~/.config/strawberry/ (config.toml, widget.cfg), voices and the widget binary ~/.local/share/strawberry/ (voices/, widget/), state and logs ~/.local/state/strawberry/ (XDG variables respected). uv build makes the wheel.

End-to-end check: scripts/check_phase1.sh (unit tests, then a headless widget against a real daemon; WIDGET=dist/strawberry-widget-… scripts/check_phase1.sh runs the same checks in the exported binary); scripts/check_tray.sh registers the tray and reads it back.

Licence

MIT, see LICENSE. Models and voices are downloaded by you under their own licences; THIRD_PARTY.md lists them and the libraries she uses.