FunButton.ai

Talk fast. Stay local. Pay less.

The fun, dev-grade voice dictation tool. Press the button, talk, watch your cleaned-up text land at the cursor. No cloud lock-in. No 800 MB Electron tax. Zero API key, zero install — cleanup runs on a bundled local model out of the box.

Status: Sprint 1 building (May 8-11, 2026). MVP target: Monday morning.

What's different

	Wispr Flow	FunButton
Bundle	800 MB Electron, 8% CPU idle	~1 GB Tauri 2 (incl. bundled LLM), ~150 MB RAM cleanup-only
Cloud	required	optional (bundled local model is the default)
Code mode	not really	first-class — spoken symbols + casing taxonomy
API key	required	not required — cleanup runs on bundled Qwen 2.5 1.5B
Linux	none	first-class (post-MVP)
Source	closed	GPLv3 desktop core
Price	$144/yr, no lifetime	$99 lifetime founder pricing (planned)

Architecture

• Tauri 2 shell — single Rust binary + WebKit webview, ~10 MB.
• cpal for audio capture (Core Audio on macOS, WASAPI on Windows, ALSA on Linux).
• rdev for global Right Option push-to-talk (modifier-only key, requires Accessibility permission).
• Groq Whisper Large v3 Turbo (~300 ms TTFT) for transcription.
• Cleanup is pluggable:
  • Embedded (default) — bundled llama.cpp server + Qwen 2.5 1.5B Instruct Q4_K_M GGUF, spawned as a child process on app launch. Zero install, zero key, runs offline.
  • Groq Llama 3.3 70B — fastest, cloud, requires GROQ_API_KEY or a FunButton license.
  • Ollama at localhost:11434 — external Ollama if you've already got it (power-user path).
  • Auto — tries embedded → ollama-external → groq, picks the first that's available.
• Code mode — auto-detected from the frontmost app (Cursor / VS Code / JetBrains / Vim / Terminal / Xcode). System prompt handles spoken symbols (open paren → (, arrow → ->, equals → =) and casing (camelCase X Y Z → xYZ, snake_case X Y Z → x_y_z, etc.).
• Text injection — clipboard write + Cmd+V via enigo, with prior clipboard restored ~1 s later.

Install

(Pre-release — until v0.1.0 .dmg exists, build from source.)

git clone https://github.com/todddickerson/funbutton.git
cd funbutton/apps/desktop/src-tauri
bash scripts/fetch-vendor-deps.sh   # one-time: pulls llama-server + Qwen 2.5 1.5B GGUF (~1 GB)
cd ..
npm install
npm run tauri build -- --target aarch64-apple-darwin
open src-tauri/target/aarch64-apple-darwin/release/bundle/macos/FunButton.app

If macOS Gatekeeper blocks the unsigned build: sudo xattr -cr /Applications/FunButton.app.

First run: macOS will ask for Microphone and Accessibility permissions. Grant both. The bundled LLM warms up in the background and you'll see a toast when it's ready (~1 s warm cache, ~10 s cold).

Note on transcription: Whisper transcription is not bundled yet. Free-tier users still need a Groq key OR a FunButton license for the speech-to-text step. The cleanup model is fully local. Bundled Whisper is on the V1.2 roadmap.

Use

1. Hold Right Option.
2. Talk.
3. Release.
4. Cleaned text appears at your cursor.

Tray menu: Settings opens the configuration window. Quit exits the app.

Develop

cd apps/desktop
npm install
GROQ_API_KEY=... npm run tauri dev

Verify the pipeline without installing

You can sanity-check that the Groq Whisper + Llama cleanup chain works on your network and key, without installing the .app:

brew install jq ffmpeg  # one time
GROQ_API_KEY=gsk_... ./scripts/test_pipeline.sh

It synthesizes a WAV via macOS say, runs the full pipeline, and prints raw vs cleaned. Pass your own phrase as the first arg.

License

GPL-3.0-or-later (desktop core). Cloud / sync / team features (when they exist) will be source-available under a separate license.

Status

See PROGRESS.md for the build heartbeat. See PRD.md for scope. See RESEARCH.md for the competitive analysis. See ~/clawd/projects/funbutton/COMPETITIVE-LANDSCAPE.md for the wedge sharpening.