Fono

Talk to your coding agent. The headline feature is an early-preview voice
loop that lets any MCP-capable coding agent — Forge, Claude Code, Cursor,
Codex CLI, Gemini CLI, and others — speak and listen through Fono. Plus a
Debian/Ubuntu install fix so the on-screen overlay shows up on first run
instead of after a manual restart.

Added

• Talk to your coding agent (early preview). Fono now ships an
  MCP server that lets any MCP-capable coding agent — Forge, Claude
  Code, Cursor, Codex CLI, Gemini CLI, and others — drive a voice
  loop through three tools: fono.speak (the agent speaks a reply),
  fono.listen (the agent asks a free-form question and gets your
  spoken answer back as text), and fono.confirm (the agent offers
  A/B/C choices and matches your spoken pick). Verified end-to-end
  against Forge and Claude Code; best-effort for the rest. Disabled
  by default, opt in with fono use mcp-server on. This is an
  early preview — the protocol, defaults, and tool surface may
  still shift before the feature graduates.
• One-command setup for your coding agent.
  fono agent-setup <name> wires everything in one shot: enables
  the MCP server, merges the right mcpServers.fono entry into
  your agent's MCP config, and appends the shared voice-mode preset
  to your project's AGENTS.md / CLAUDE.md. Idempotent, supports
  --dry-run, and --list shows every registered agent. After
  setup, launch your agent the normal way and it can speak and
  listen.
• Voice-friendly overlay while the agent is talking with you.
  When the agent calls fono.listen the same overlay you see for
  F7 dictation pops up — waveform/transcript while you speak, a
  PONDERING animation while it waits — so you always know whether
  Fono is listening. The overlay is scoped strictly to the
  microphone-open phase (not while the agent is speaking its
  prompt) and is skipped when a regular Fono daemon is already
  running so a daemon-paired environment never double-paints.
• Background-speech filter. When the agent asks a question,
  Fono now filters out chatter that doesn't look like an answer —
  radio, TV, a side conversation in the room, or the agent's own
  prompt echoing back through the speakers. Tunable via
  [mcp].relevance_filter ("off" | "heuristic" | "llm", default
  "heuristic") and [mcp].relevance_max_rejections (default
  2). The optional "llm" mode uses the configured polish
  backend as a one-shot classifier with a 1.5 s timeout; on
  timeout or parse failure it fails open and accepts the
  utterance. Each rejection flashes a dim grey Ignoring badge in
  the overlay so you can see that Fono heard you but is still
  waiting for a real answer.
• Tray icon turns amber while the agent is in a voice turn.
  Same colour Fono uses while STT or polish is running, so you can
  tell at a glance that a fono.listen / fono.speak /
  fono.confirm call is in flight. The previous tray state is
  restored when the call ends; nested spans (a prompt that speaks
  before listening) keep the icon steady. No configuration needed.
• fono speak --stream — reads stdin line by line, segments
  into sentences, strips markdown, and speaks each sentence
  through the configured TTS backend. Backpressure prevents a
  fast producer from outrunning playback; SIGINT flushes cleanly.
  Pipe-friendly: echo "Hello. World." | fono speak --stream.
• fono use mcp-server on|off — toggle [mcp.server].enabled
  without editing config by hand.
• fono doctor "Coding agents" section — reports whether the
  MCP server is enabled, the last MCP handshake timestamp, the
  advertised tools, and the last tool-call result.
• Tray "MCP server" submenu (visible only when enabled) —
  enable/disable toggle, last-connected timestamp, per-tool
  enable/disable rows.
• Docs. New docs/coding-agents.md integration guide covering
  Forge, Claude Code and Cursor (verified) plus Codex CLI, Gemini
  CLI, Cline, Continue, Windsurf and Goose (best-effort), with an
  "Adding your own agent" section. Shared voice-mode preset at
  assets/agent-presets/voice.md.
• ADR 0030 — records the Fono-as-MCP-server decision, the
  three-tool surface, the agent-agnostic design principle, and the
  agents.toml registry.

Changed

• MCP listen default silence is now 10 s (was 2 s) so an
  agent turn can pause for thought without being cut off
  mid-sentence.
• MCP listen default max_seconds lowered from 60 s to 45 s —
  combined with the multi-utterance relevance loop this gives a
  responsive turn-taking budget without stranding you.

Fixed

• Overlay now appears on first run on Debian/Ubuntu desktops.
  Installing via curl https://fono.page/install | sh silently
  skipped the post-install prompt that offers to add
  libxkbcommon-x11 and xdotool, so the on-screen recording
  overlay fell back to noop and only appeared after a manual
  restart. The prompt now reads from /dev/tty directly so it
  survives curl|sh + sudo PTY allocation, and the background
  daemon spawn also reconstructs DISPLAY and XAUTHORITY when
  sudo strips them. Server installs are unaffected — they never
  pull X11 libraries.

Full Changelog: https://github.com/bogdanr/fono/compare/v0.8.2...v0.9.0