t2t

t2t logo

Push-to-talk dictation for macOS. Hold Fn to type. Hold Fn + Ctrl to talk to a local Pi agent.

t2t is a small menu-bar app that turns a keyboard hold into a complete voice workflow. Dictation runs through Whisper on your Mac. Agent mode hands the transcript to the pi CLI you already use, with your configured model and MCP tools.

Features at a glance

Feature What happens Available by default
Local dictation Hold Fn, speak, and release. Whisper transcribes the recording locally. Yes
Paste into the focused field The transcription is inserted into the text field that was focused when the turn started. Yes
Agent mode Hold Fn + Ctrl to send the transcript to the local pi CLI. Requires pi on PATH
Spoken agent replies T2T can read the agent response aloud with macOS text-to-speech. Configurable in Settings
MCP tools Add, edit, remove, enable, disable, and inspect MCP servers from Settings. Agent turns can use the enabled servers. Configurable in Settings
MCP directory Browse example MCP servers for files, databases, GitHub, browsers, and documentation. Included on t2t.now
Local history Keep transcription and agent-turn history on the Mac for later inspection. Yes
Live agent history Settings → History shows agent activity and streamed response text while a turn is running. Yes
Two-mode indicator Transparent, click-through indicator bars show recording, processing, and speaking state. Yes
Multiple displays Indicator surfaces are placed on connected displays and stale display surfaces are removed after disconnects. Yes
Cancellation Press Escape to cancel an active processing turn. Yes
Screen context Agent turns can include a temporary screenshot as visual context; the temporary file is deleted after the turn. Configurable by the agent flow
Privacy controls Audio transcription is local. Credentials stay in your local Pi configuration; T2T does not store them in the repository. Yes

Download

Download for macOS

View releases on GitHub

T2T is currently ad-hoc signed for local development. A rebuilt app can receive a new macOS identity, so macOS may ask you to approve Accessibility, Microphone, and Input Monitoring again.

Requirements

  • macOS
  • Accessibility permission, for the global Fn shortcut and focused-field insertion
  • Microphone permission, for recording
  • pi or a compatible pi CLI on PATH for agent mode
  • A Pi provider and model configured locally

First run

  1. Download or build the app.
  2. Open T2T and grant Accessibility and Microphone access when macOS asks.
  3. Hold Fn in any editable field and speak.
  4. For agent mode, install pi, configure a provider and model, then hold Fn + Ctrl.
  5. Open Settings to configure voice output, Pi options, MCP servers, and history.

The first local transcription run uses the Whisper model at ~/.cache/whisper/ggml-base.en.bin. T2T downloads or initializes that model through its local Whisper path.

Pi and model configuration

T2T delegates agent turns to the locally installed pi CLI. It does not embed a hosted chat client or copy Pi credentials into the app. The current public defaults are:

  • Provider: cloudflare-ai-gateway
  • Model: gpt-5.6-luna
  • Thinking: medium
  • Session mode: --no-session

Override these values in Settings or with environment variables:

export T2T_PI_BINARY=pi
export T2T_PI_PROVIDER=cloudflare-ai-gateway
export T2T_PI_MODEL=gpt-5.6-luna
export T2T_PI_THINKING=medium

The provider, model, thinking level, and optional CA bundle can also be stored in T2T's local Pi-agent settings.

MCP servers

MCP is optional. In Settings, you can manage the servers that the Pi agent is allowed to discover and use for a turn. T2T preserves unrelated MCP configuration fields when it updates the server list and treats a server as available only when its runtime configuration explicitly enables it.

Examples include:

  • Filesystem access
  • Git and GitHub
  • PostgreSQL and other databases
  • Browser and developer-tools automation
  • Documentation and search services

MCP tools are not blanket authorization. A server can expose a capability, but the current voice request still needs to authorize an external or state-changing action.

Local data and history

  • Transcriptions and agent-turn summaries are stored locally.
  • Pi credentials remain in Pi's local configuration.
  • Logs: ~/Library/Logs/t2t.log
  • Whisper model: ~/.cache/whisper/ggml-base.en.bin
  • Temporary screenshots used for an agent turn are deleted after use.

Developer setup

Validate the desktop app

cd /Users/jcoeyman/cloudflare/t2t/desktop
bun install
bun run check
cargo test

Run desktop development mode

cd /Users/jcoeyman/cloudflare/t2t/desktop
bunx tauri dev

Development builds have a different macOS identity from /Applications/t2t.app, so permissions granted to one copy may not apply to the other. Run only one T2T instance at a time.

Build and install the canonical app

cd /Users/jcoeyman/cloudflare/t2t/desktop
bun run macos:install
open -a /Applications/t2t.app

Do not launch the copy inside desktop/target for normal use. The install script builds the release app and replaces /Applications/t2t.app.

Repair macOS permissions after a rebuild

# Stop only T2T, if it is running.
pkill -f '/Applications/t2t.app/Contents/MacOS/t2t' 2>/dev/null || true

# Reset only T2T's grants. macOS will ask for approval again.
tccutil reset Accessibility com.t2t.desktop
tccutil reset Microphone com.t2t.desktop
tccutil reset ListenEvent com.t2t.desktop

open 'x-apple.systempreferences:com.apple.preference.security?Privacy_Accessibility'
open 'x-apple.systempreferences:com.apple.preference.security?Privacy_Microphone'
open 'x-apple.systempreferences:com.apple.preference.security?Privacy_ListenEvent'
open -a /Applications/t2t.app

Verify the installed app

ps -axo pid,args | grep '[t]2t.app/Contents/MacOS/t2t'
ls -ld /Applications/t2t.app
cd /Users/jcoeyman/cloudflare/t2t
git diff --check

Build without installing when you only need an artifact:

cd /Users/jcoeyman/cloudflare/t2t/desktop
bun run build

Website development and deployment

The public site lives in web/ and is deployed to t2t.now with Alchemy and Cloudflare.

cd /Users/jcoeyman/cloudflare/t2t/web
bun install
bun run check
bun run build
bun run deploy

bun run deploy uses the configured Cloudflare credentials and Alchemy state. It may ask for the deployment password or other environment variables when they are not present locally.

Tech stack

  • Tauri 2 and Rust
  • Svelte 5 and SvelteKit
  • whisper-rs for local transcription
  • Local pi CLI bridge for agent mode
  • MCP server management and discovery
  • Cloudflare Workers, D1, Durable Objects, and Workers AI for the website
  • macOS say for spoken agent replies

License

MIT