acoyfellow/t2t

t2t is a macOS menu-bar app that turns your keyboard into a voice workflow. Hold the Fn key, speak, and your words are transcribed and typed into whatever field you were focused on. Hold Fn plus Ctrl and your spoken text goes to a local AI agent that can use tools to act on your request. The transcription happens entirely on your Mac using Whisper, so no audio leaves your machine. The first time you use it, the app downloads a local speech model. For agent mode, it hands your transcript to a separate command-line tool called pi that you install and configure yourself. It doesn't embed a chat client or store your credentials. MCP servers (think of them as plugins that give the agent capabilities like filesystem access, GitHub integration, or database queries) can be managed from the settings panel, letting you choose which tools the agent can reach during a voice turn. Anyone who dictates notes, emails, or code would appreciate the push-to-talk simplicity. But the real value is the agent mode for people who want to speak a request like "search my GitHub repos for X" or "look up this in my database" and get a spoken answer back. The app can read agent responses aloud using macOS text-to-speech, and it keeps a local history of both transcriptions and agent interactions. A few things stand out. Privacy is a clear priority: audio stays local, credentials live in your existing pi configuration, and temporary screenshots used for visual context are deleted after each turn. The app also shows transparent on-screen indicators for recording, processing, and speaking states across multiple displays. One limitation worth noting is that it currently runs only on macOS, and agent mode requires installing the pi CLI separately.