Voice input for AI coding agents¶
Short answer: hold a key, speak your prompt, let go. YazSes types it into whatever has focus: a terminal running Claude Code, Codex or Aider, a chat box in a desktop app, or a browser tab. Speech recognition runs on your own CPU, so the prompt reaches the AI tool the same way typed text does, and no transcription service ever hears it.
Linux is the best-supported platform. On macOS and Windows YazSes is a preview.
Why agent prompts are hard for ordinary dictation¶
A good prompt for a coding agent is long, and you think while you say it. Two common designs get that wrong:
- Voice modes that stop on silence treat your pause for thought as the end of the prompt and cut it off. YazSes uses hold-to-talk: it listens for as long as you hold the key, pauses included, and stops only when you let go.
- Cloud dictation sends your audio to a server before the text reaches the agent. That audio can contain code, customer names and unreleased plans. YazSes transcribes entirely offline.
The basic flow¶
- Focus the agent's input: the terminal where
claudeorcodexis running, or a chat box. - Hold your hotkey (Right Alt on Linux by default), speak the prompt, and release.
- Read the text, fix anything that is wrong, and press Enter yourself.
This needs no setup on the agent's side, and it works over SSH: if the agent runs on a remote machine in your local terminal, the keystrokes travel down your SSH session like typed ones. See over SSH and in Remote-SSH editors.
Review before it types: staged dictation¶
A terminal agent acts on Enter, and one misheard word can change an instruction. Staged dictation holds what you say in a buffer instead of typing it:
Dictate one burst or several, then:
yazses staged status # what is waiting
yazses staged commit # type it into the focused window
yazses staged undo # drop the last burst
Bind yazses staged commit to a key if you use it often. Saying "commit that" can be turned on with [staged] spoken_commit. It is off by default because, when recognition goes wrong, the commit phrase is exactly what can be misheard.
Hand a prompt to the Claude or Codex app: send to¶
If the agent is the Claude desktop app or the Codex app, YazSes can open it with your prompt already in the box, so you do not have to focus its input before you speak. It fills the box and stops there: you read the prompt and press Enter yourself. It is off by default and has not yet been tried against the real desktop apps.
Turn it on, then restart the daemon so it listens for the phrases:
That writes enabled = true under [sendto] in config.toml, which you can also set by hand.
Then dictate, review, and send it over:
- With staged mode on, dictate the prompt into the buffer.
- Say "send to Claude", "send to Claude Code" or "send to Codex" as a whole utterance, or run
yazses send-to claude(orclaude-code,codex). - The app opens a new conversation with the prompt filled in. Read it, edit it, and press Enter there.
Without staged mode, the phrase hands over the last burst YazSes typed. To send text from a script or a key binding, give it directly: yazses send-to codex "explain the failing test".
A few limits are deliberate. Only those three apps can be targets, through their own claude:// and codex:// links; there is no web fallback, because opening a web page would send the text to a server before you had read it. A prompt over 14,000 characters is refused rather than cut, and stays in the staged buffer. The log records which app and how many characters, never the words. Details are in the CLI reference.
"Never sent" is something YazSes checks against each real app, not something it takes from the apps' documentation. yazses features info sendto names the app versions it has been verified on; today it says it has not been verified against any. If you have the Claude or Codex app on Linux, you can help: the steps are in the validation record of ADR-031. Use a made-up prompt, never a real one, and post what you saw on issue #637. An app that ever sends the prompt by itself is removed as a target in the next release.
Technical vocabulary¶
Agents need the names in your codebase spelled correctly. Add the project's identifiers, library names and acronyms to your personal vocabulary. For symbols and code itself, see coding by voice.
Let an agent use YazSes: the MCP server¶
YazSes also works in the other direction: an agent can call it. yazses mcp-server is an MCP server over stdin/stdout, with no network port. Add it to an MCP client's configuration:
It offers transcribe, which turns an audio or video file into text on your machine. With [mcp] ask_human = true it also offers ask_human: an agent that needs a decision only you can make asks you out loud, and gets your spoken answer back instead of having it typed into your editor. A budget limits how often agents may interrupt you. Details are in the CLI reference, and setup for each client is in Connect YazSes to an MCP client.
Not released yet¶
One more piece is built but not yet in a release:
- One-click MCP: install the MCP server from a client's extension screen or the MCP registry, and an
ask_humanthat stops listening when you stop talking (ADR-032).
The bundle and the registry entry are built but have not shipped in a release yet, and the registry listing waits for a manual publish. Send to, the other bridge, is described above (ADR-031). The roadmap explains why this is where YazSes is going next.
What YazSes will not do¶
It will not send anything on your behalf, route what you say to an online assistant, or listen all the time for a wake word. You press Enter, and you choose the tool.