Install one-liner

curl -fsSL https://fono.page/install | sh

The script picks the right binary for your CPU (and switches to the Vulkan-GPU build if your machine has one), runs sudo fono install to place it on $PATH, starts the daemon, and opens the fono setup wizard in the same terminal.

Different styles

While you're speaking, a small overlay shows what the microphone is hearing. Four styles ship: bars, oscilloscope, fft, heatmap. Switch via the tray (Preferences → Waveform style) or set [overlay].style in ~/.config/fono/config.toml.

What Fono does

• Dictation, push-to-talk or toggle. Tap F7 to toggle recording; hold F7 for push-to-talk. The same key works either way — the press duration decides.
• Voice assistant on F8 Talk to Ollama, OpenAI, Anthropic, Groq, Cerebras, or OpenRouter; the reply is streamed sentence-by-sentence into TTS so audio starts before the model has finished thinking.
• Talk to your coding agent (early preview). Drive Forge, Claude Code, Cursor, Codex CLI, Gemini CLI — any MCP-capable agent — entirely by voice. fono agent-setup <name> wires it up in one shot; after that the agent speaks its replies, listens for your follow-ups, and offers A/B/C choices you pick with your voice. See docs/coding-agents.md.
• Lands in any X11 or Wayland window. Fono types straight into the focused window and mirrors to the clipboard as a safety net. Hotkeys register through the Wayland portal where it's available, with automatic fallbacks for GNOME 46 and X11 — see docs/wayland.md for the per-compositor story.
• Local or cloud speech-to-text. Whisper runs on your machine by default. Or switch to Groq, OpenAI, or Deepgram with one command (fono use stt …).
• Local or cloud text-to-speach. For local you can use Wyoming-piper. More options are on the roadmap. Or switch to various cloud providers if you want most naturally sounding voices.
• Automatic model selection. The first-run wizard probes your CPU and GPU, then picks the heaviest local Whisper model that runs better than real time on your hardware — no manual tuning. The decision matrix was engineered here with older and new machines accross multiple days of benchmarking.
• Optional cleanup pass. A small LLM can tidy up the transcript before it's injected — locally with llama.cpp, or via Cerebras / Groq / OpenAI / OpenRouter / Anthropic / Ollama.
• Visualisation overlay during recording. Bars, oscilloscope, FFT, or heatmap. Live-dictation mode adds a small VU bar.
• Optional GPU acceleration. fono update probes your host for Vulkan and pulls the matching CPU or Vulkan build automatically.
• LAN-friendly. Speaks the Wyoming protocol as both client and server, so Fono can route through (or host for) a Home Assistant satellite or another Fono on the network. mDNS finds peers automatically.
• Two small builds CPU/GPU ~22/60 MB No Electron, no Node, no Python, no WebKit. Four glibc dependencies.

First run

sudo fono install installs the files and starts the setup wizard.

Default hotkeys are F7 (dictation) and F8 (voice assistant). Both keys auto-detect how you press them: a quick tap toggles recording on (tap again to stop); holding for more than a second turns the key into push-to-talk and recording ends on release. Escape cancels a recording or shuts up an assistant reply.

The setup wizard hot-reloads the running daemon when it finishes, so you don't need to restart anything. Reconfigure with fono setup.

Switching providers

fono setup asks for a primary cloud provider. With OpenAI or Groq, a single API key covers STT, cleanup, the assistant, and TTS. Narrower providers (Anthropic, Cerebras, OpenRouter) cover what they offer; the wizard only prompts for follow-on keys if you opt in to capabilities they don't cover.

fono use cloud groq           # paired preset (Groq STT + Groq LLM)
fono use stt openai           # change just STT
fono use tts cartesia         # swap TTS backend
fono use local                # back to whisper-local + skip polish

Keys live in ~/.config/fono/secrets.toml:

fono keys add GROQ_API_KEY    # paste at the prompt
fono keys check               # reachability probe per stored key

TTS works with OpenAI, Groq, OpenRouter (Kokoro), Cartesia, Deepgram, and any Wyoming server you have on the LAN.

Other ways to install

• Distro packages. .deb, .pkg.tar.zst, and .txz files are built by CI and attached to each release, but they are not regularly tested — they may work, please file an issue if they don't.
• macOS and Windows. Planned, not shipping.

Privacy

Local-first. Nothing leaves your machine unless you pick a cloud provider.

Documentation

• Documentation index — the full map.
• Install — one-liner, manual install, server mode, updating.
• Quickstart — first dictation, common follow-ups.
• Configuration — every key in config.toml.
• Provider matrix — STT, polish, assistant, and TTS endpoints.
• Live dictation — streaming overlay, latency budget.
• Troubleshooting — symptom-first recipes.
• Roadmap — in progress, planned, shipped.
• Homepage: fono.page.

Status

Linux-first; used daily by the maintainer. Rough edges exist — issues and patches are welcome. See ROADMAP for what's next.

Contributing

Pull requests welcome. See CONTRIBUTING.md for the workflow (DCO sign-off required).

License

GPL-3.0-only. See LICENSE.