Fono

Fono can now share the AI model you already have configured with everything
else on your machine and LAN: turn on one switch and Fono answers requests
on a local, OpenAI- and Ollama-compatible HTTP API — so your editor, Open
WebUI, llm, LangChain, and Home Assistant's Ollama conversation agent can
all talk to Fono's model with no extra setup.

Added

• Serve your local model over a standard API ([server.llm], default
  off). Flip it on — from the config file or the tray (Servers → Local
  LLM server) — and Fono exposes whatever assistant you have configured on
  a local HTTP endpoint that speaks both the OpenAI
  (/v1/chat/completions, /v1/models) and Ollama-native (/api/chat,
  /api/tags) formats, on Ollama's usual port 11434 so existing clients
  connect unchanged. Editors, Open WebUI, llm, LangChain, and Home
  Assistant's Ollama conversation agent can use Fono as their model backend
  out of the box. It binds to loopback by default, takes an optional bearer
  token before you expose it on the LAN, hot-reloads from the tray without a
  restart, and follows a backend swap (fono use assistant …) on the next
  request. fono doctor reports whether it's running and which model it
  serves, and Fono advertises it over mDNS so other machines discover it.
  Design and rationale in
  ADR 0036.
• Full cloud fidelity via pass-through. When the model you serve is an
  OpenAI-compatible cloud provider (OpenAI, Gemini, Groq, Cerebras,
  OpenRouter), Fono forwards the request straight to the provider — injecting
  your stored key on the way out and keeping it on the machine — and streams
  the reply back unchanged. Every model the provider offers, plus
  tool/function-calling, vision, and JSON mode, work exactly as if you called
  the provider directly, with nothing to configure. This is what lets Home
  Assistant drive smart-home devices through a cloud model behind Fono.
• A one-line request log. Run the daemon with --debug (or
  FONO_LOG=fono::llm::server=debug) and Fono prints one tidy line per
  request: the endpoint and status, whether it was served locally or passed
  through to a cloud provider, the model, timing (time-to-first-token and
  total), an output-token count and throughput where available, and which app
  made the call (from its User-Agent) so you can tell clients apart on a
  shared port. Prompt and reply content are never logged.

Changed

• Realtime assistants keep working over the API automatically. If your
  assistant is a realtime speech-to-speech model (e.g. Gemini Live) that a
  text chat API can't expose, Fono serves the same provider's fast text
  model instead (for Gemini, gemini-flash-lite-latest), reusing the same
  key — so you keep Gemini Live for voice and get a cheap, fast, smart text
  model on the API at the same time, with zero extra configuration. An
  optional [server.llm].model pins any specific model.

Full Changelog: https://github.com/bogdanr/fono/compare/v0.12.0...v0.13.0