t2t – Voice-to-Text with MCP Agent Mode on macOS

TL;DR

TL;DR: t2t is a macOS menu-bar app that turns keyboard holds into a complete voice workflow — local Whisper transcription on fn, MCP agent control on fn+ctrl, built with Tauri, Svelte, and Rust.

Source and Accuracy Notes

⚠️ This section is MANDATORY. All links must be verified from actual source, not guessed.

• Project page: t2t.now
• Source repository: github.com/acoyfellow/t2t
• License: MIT (verified via LICENSE file in repo)
• HN launch thread: news.ycombinator.com/item?id=show hn — 4 points, Show HN posted by author
• Topics verified: mcp, whisper, local-first, accessibility, clipboard, push-to-talk, rust, tauri, svelte, sveltekit

What Is t2t?

t2t (Talk to Type) is a menu-bar app for macOS that intercepts the fn key as a voice input trigger. Hold fn in any text field and speak — Whisper transcribes locally and inserts the result. Hold fn+ctrl and the transcript goes to the locally configured pi CLI, which routes it through an AI provider to any enabled MCP server.

The distinction from a typical voice dictation tool is the agent mode: you are not just dictating text, you are running a full voice-driven agent loop with access to MCP tools for files, databases, Git, browser automation, and more.

Setup Workflow

Step 1: Download and install

Download the macOS binary from the project page:

open https://t2t.now

Or grab a release directly from GitHub:

open https://github.com/acoyfellow/t2t/releases

The app is ad-hoc signed. On first launch macOS will ask you to approve Accessibility (for global shortcut detection and field insertion) and Microphone permissions.

Step 2: Grant permissions

Open System Settings → Privacy & Security → Accessibility and ensure t2t is listed and enabled. The same for Microphone under Privacy & Security.

Step 3: Dictate with fn

Open any text field — a browser search bar, a code editor, a Notes document. Hold fn, speak, release. The transcription appears in the field you were focused on.

The first run downloads the Whisper model to ~/.cache/whisper/ggml-base.en.bin.

Step 4: Configure agent mode (optional)

Agent mode requires the pi CLI on your PATH. Install it and configure a provider and model:

# Install pi (from github.com/badlogic/pi-mono)
# Then configure in Settings or via env vars:
export T2T_PI_PROVIDER=cloudflare-ai-gateway
export T2T_PI_MODEL=gpt-5.6-luna
export T2T_PI_THINKING=medium

Or set them directly in t2t’s Settings UI under Pi options. The default provider is Cloudflare AI Gateway with gpt-5.6-luna.

Step 5: Add MCP servers

In Settings → MCP, add servers for the tools you want the agent to access. The t2t website includes an MCP directory with example servers for:

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

Enable only the servers you trust — MCP authorization is per-capability, not blanket.

Deeper Analysis

Architecture

t2t is built with Tauri (Rust backend) + SvelteKit (UI). The fn global shortcut is handled by Tauri’s app-level event system. Audio recording uses the macOS microphone permission, and Whisper runs entirely on-device for transcription.

Agent mode delegates to the pi CLI, which is a separate installed binary. t2t does not embed credentials or a hosted chat client — it pipes the voice transcript to pi and reads back the streamed response.

Privacy model

All audio transcription is local. Credentials stay in pi’s local configuration. The app stores transcriptions and agent-turn history locally at ~/Library/Logs/t2t.log. Temporary screenshots captured for agent context are deleted after each turn.

Two-mode input

| Shortcut | Function | |---|---| | fn (hold) | Record audio → Whisper transcription → insert into focused field | | fn+ctrl (hold) | Record audio → Whisper transcription → send to pi agent → TTS readback |

An on-screen indicator shows recording, processing, and speaking state. Escape cancels an active turn.

Multi-display support

Indicator bars appear on all connected displays. If you disconnect a display, stale indicator surfaces are removed automatically.

Practical Evaluation Checklist

• Local Whisper transcription — audio never leaves the machine
• fn works system-wide — any focused text field accepts input
• Agent mode with MCP — voice-driven access to files, Git, DBs, browser
• Cross-platform — macOS primary; Linux and Windows listed in topics
• Open source MIT — no vendor lock-in for the core app
• Configurable TTS — agent responses can be read aloud via macOS text-to-speech
• Local history — full transcript and agent-turn log retained on-device

Security Notes

• Microphone and Accessibility permissions are required. Review what you grant.
• MCP servers are configurable per-capability. Do not enable servers you do not trust.
• Credentials are stored in pi’s local configuration, not in the t2t app itself.
• The pi agent runs locally on your Mac — it is not a hosted service.
• Temporary screenshots captured for agent context are deleted after each turn.

FAQ

Q: Does t2t work on Linux or Windows? A: The project page targets macOS as the primary platform. The GitHub repository topics include linux and windows, indicating cross-platform support may be available or planned, but the primary download and documentation focus on macOS.

Q: What Whisper model does it use? A: The default model is ggml-base.en.bin, downloaded on first run to ~/.cache/whisper/. You can swap this for a different Whisper model if needed.

Q: What AI provider does agent mode use? A: Agent mode delegates to the pi CLI. The default provider is Cloudflare AI Gateway using gpt-5.6-luna, but you can configure any provider and model that pi supports via environment variables or the Settings UI.

Q: Can I use t2t without an MCP server? A: Yes. Basic dictation mode (hold fn) works without pi or any MCP server — just Whisper transcription inserted into the focused field.

Q: How is t2t different from macOS native dictation? A: macOS dictation (hold fn twice) is limited to transcription. t2t adds agent mode (via pi + MCP), spoken agent responses via TTS, local history of all turns, and a configurable shortcut (fn+ctrl) for the agent loop.

Conclusion

t2t is a well-scoped voice workflow tool that solves a real ergonomic problem: typing interrupts flow. The fn hold metaphor is tactile and context-aware — you are never unsure which mode you are in. Adding MCP agent access on fn+ctrl extends it from a dictation tool into a hands-free coding assistant.

It is MIT licensed, runs Whisper locally, and delegates AI to a self-managed pi configuration — no built-in cloud dependency. If you spend significant time in a terminal or code editor, t2t is worth trying.