flow

Flow — dictation, the Wispr-Flow way, as a native Switchboard app

Hold a key, talk, and clean text lands at your cursor. The twist: Flow ships no AI. It's a thin shell that borrows your machine's intelligence through the Switchboard daemon — proving the "beyond web apps" path. No browser, no API key, no bundled model.

  mic  ──▶  daemon claude_transcribe  ──▶  daemon claude_complete  ──▶  your cursor
           (local whisper, on-device)      (local model if present,      (paste)
                                            else your own Claude)

100% on-device (optional): cleanup — stripping um/uh and fixing punctuation — is pure text, so it runs on a small LOCAL model when one is available. Install Ollama and pull a tiny model (ollama pull llama3.2:1b or qwen2.5:0.5b); Flow auto-picks it and the pill reads "on-device." With no local model it falls back to your Claude, and if that's unavailable it ships the raw transcript — dictation never fails. Point cleanupModel in ~/.flow/config.json at a specific model to pin it.

Flow authenticates as its own least-privilege principal, native@ai.thelastprompt.flow, over the daemon's native listener (a per-app token, never the browser's pairing key). It can only ever act as itself.

What it demonstrates

• A local app talks straight to the daemon — the extension/browser is not involved.
• Speech-to-text runs on-device (whisper), reached via the gated claude_transcribe verb.
• The "Flow magic" — stripping um / uh / you know into clean prose — runs on your own Claude via claude_complete, gated by the same broker as every web wrapp.
• The app holds no key and no model; setup grants it a scoped, revocable identity.

Requirements (all local, no cloud)

Need	Install
ffmpeg (mic capture)	brew install ffmpeg
whisper (on-device STT)	pip install openai-whisper
macOS say, osascript, pbcopy	built in

Build the daemon once: npm run -w @relay/sidekick build.

Run it

# No mic needed — macOS `say` generates rambly speech, runs the whole pipeline, prints before/after:
node examples/flow/flow.mjs demo "um so I think we should uh ship on friday you know"

# The real thing — push-to-talk mic, pastes cleaned text at your cursor:
node examples/flow/flow.mjs run

run is push-to-talk: ENTER to start recording, speak, ENTER to stop. The cleaned text is copied to your clipboard and auto-pasted into the frontmost app (grant your terminal Accessibility for the auto-paste; otherwise it's on the clipboard to ⌘V yourself).

How it's wired

• Setup (once): flow.mjs reads the daemon's pairing token and calls the registerNativeApp control action — the native equivalent of the connect-consent. That mints Flow's per-app token (saved 0600 in ~/.flow/token.json) and a grant for the local capabilities + your online models.
• Every run after: Flow uses only its per-app token on the native socket. Least privilege.
• STT backend: the daemon runs whatever RELAY_STT_CMD points at; Flow points it at whisper-stt.mjs (audio path in → transcript out). Swap in Apple's Speech framework or a whisper.cpp server by changing that one env var — nothing else moves.

For a self-contained demo, Flow manages its own daemon instance (private ~/.flow/relay, ports 8795/8796) so it never touches your real ~/.relay. In production it would attach to the menubar daemon once that ships STT configuration.

Config (env)

Var	Default	Meaning
FLOW_MIC	:0	avfoundation audio device (ffmpeg -f avfoundation -list_devices true -i "" to list)
FLOW_WHISPER_MODEL	tiny	whisper model (tiny/base/small…)
FLOW_PORT / FLOW_NATIVE_PORT	8795 / 8796	Flow's private daemon ports
FLOW_HOME	~/.flow	Flow's state (token + private daemon dir)