macOS · Apple Silicon · Local-first
Hold a hotkey and talk. Let go, and it becomes text, an answer to your question, or a task handed to an agent.
Paste this into Terminal
curl -fsSL https://opentype-site.vercel.app/install | zsh
Or hand this sentence to your AI assistant
Please fetch the instructions at https://opentype-site.vercel.app/agent and follow them to install OpenType on this Mac for me
New in 2.1.2: what changed.
One continuous take, start to finish — has sound.
Has soundSay something like “agent mode” mid-recording and this one takes that mode instead.
Straight transcription, with no model in the loop for recognition — what it heard is what you get. Direct writes straight into the field you were typing in; Review stops in an editable panel first, so you can fix a misheard word before it lands.
Speak a question, get the answer in the overlay. It will search the web when that genuinely helps, and cite what it used. The answer stays in the overlay and on your clipboard — it never types itself into whatever you had open.
Speak a task and hand it over. It has real hands — shell, Python, reading files, searching directories, the web — plus any MCP tools you connect. What comes back is a draft: it never presses send for you.
Agent mode runs real shell commands, Python and file edits on your Mac — no sandbox, and no confirmation prompt for most of them. What it hands back is always a draft: copied to your clipboard, never sent, published or executed on your behalf.
MLX-Whisper transcribes offline on your Apple Silicon by default. The audio goes to no server and is deleted once it has been read. You can point it at a remote endpoint instead, if you want to.
Bring Anthropic or anything speaking the OpenAI protocol — DeepSeek, OpenAI, your own server. Test the connection, then pick from the model list it fetches, instead of typing a model name and hoping.
Tell it “remember that when I say PayPal I mean the company”. Your terms and the facts about you stay on this machine, and Ask and Agent draw on them when they answer.
Audio is transcribed on-device by default and deleted after; the only thing that ever leaves is text, and only to the model provider you configured, sent only by Ask and Agent. But what gets sent is more than just what you just asked — the last 10 turns go along as context, dictation included, and there's no switch to turn that off. Everything else, history, memory, your API key, stays in a local file under your own account, with no cloud sync.
The installer asks where speech recognition should run — press Return for the on-device default if you are not sure. Plain dictation needs no API key and no network.
Paste this into Terminal
curl -fsSL https://opentype-site.vercel.app/install | zsh
Or hand this sentence to your AI assistant
Please fetch the instructions at https://opentype-site.vercel.app/agent and follow them to install OpenType on this Mac for me