Speak. Keep the words. Send the work.

Hold to talk. Let go to finish.

Use a configurable global shortcut to hold for dictation or tap to toggle. On Linux and macOS the text goes in through the clipboard. Windows is coming soon. The rail lets you change the shortcut and shows detected conflicts.

Choose where your audio goes.

The picker lists 11 transcription pipelines, including local models and cloud services. Normal users get Gemma Local, Nemotron Local and Whisper Offline. Groq, OpenAI Realtime and Cohere need matching user keys. Cloud lanes ask you to acknowledge the vendor's terms and audio route.

Eleven pipelines. Not eleven free cloud accounts.

Dictate without sending audio to a cloud.

Gemma 4 runs fully offline. Nemotron 3.5 uses a 0.6B streaming model with 560 ms chunks and falls back from CUDA to CPU. Chunk size describes how it processes audio, not how long you will wait for a finished sentence.

Give specialist words a fair hearing.

Choose from 473 domain dictionaries with phonetic and fuzzy matching. Custom hotwords and abbreviations learn from usage, and your terms take precedence.

Those figures come from our own notes. A measurement on your machine may differ.

See what heard you, line by line.

A 16-bar equalizer follows real microphone energy. Streaming pipelines show partial words; batch pipelines finish after stop. Every transcript line names its answering engine. A fallback keeps the requested engine visible too, and provider filters appear only for engines present in the transcript.

Send words and pictures to the coding session.

Attach a real C0CAP capture to a dictation line. Inject that line or the full transcript into the hosting C0CODE split's composer with image previews. A merged chunk allows five captures. One composer injection allows six images. Over-limit images produce a warning.

Let finalized speech reach the board.

Turn on Send speech to the board for a session. It starts off. Finalized phrases can draw deterministic diagrams on C0CAP; free-form requests go to the Core Agent. The diagram parser uses no network or model call. Your chosen speech pipeline still has its own requirements.

Keep a spoken phrase for text you repeat.

Create a snippet with exact multiline expansion and alternate phrases. The manager reports collisions with T3 user commands and names the winner. Pronunciation recording uses local Whisper, stores the recognized text alias, and discards the audio. It matches phrases, not your voiceprint.