Voice input for AI agents
Dictation for people whose day is mostly prompts. The words arrive in the agent’s input box, not in the agent — including over SSH, and including in the agents that have no voice feature at all.
Can you dictate prompts to an AI agent?
Yes, with dictation that types rather than sends. VibeVoice puts the transcript into whichever text field has focus — a terminal running an agent, a browser chat box, an editor panel — and stops there. You read it, fix a term, add what you forgot, and send when you mean to.
It works the same in agents with no voice feature of their own, and in a terminal connected over SSH, because nothing is integrated: the client types into the local window the way a keyboard does.
Built-in voice usually stops where the work is
Claude Code has the most thoroughly documented voice feature of any coding agent, so it is the fair one to quote. All of this is from its own documentation.[1]
Not over SSH
Dictation needs a local microphone and does not work over SSH or in remote VS Code sessions.[1] A great many agent sessions are exactly that.
Tap mode stops at two minutes
Recording ends after fifteen seconds of silence or two minutes in total.[1] A pause to think reads to a timer as silence.
What each agent’s own documentation says
Two of them do editable dictation natively and are in the table for that reason.
| Agent | Native voice | What its docs describe |
|---|---|---|
| Copilot CLI | Yes | Inserts the transcript at the cursor to edit before submitting. Needs a runtime and model download, and lists English and Spanish.[2] |
| Aider | Yes | /voice records inside the chat and writes the speech as if typed — within Aider’s own session.[3] |
| Claude Code | With limits | Local mic only, not over SSH or remote VS Code; in tap mode, 15s silence and a two-minute cap, and it auto-submits.[1] |
| Codex | In the app | Voice lives in the ChatGPT desktop app and its session model: a chat must start in voice mode, one at a time, on an eligible plan.[4] |
| Gemini CLI | Proposed | An open request is answered with OS dictation or a third-party setup;[5] another asks for cursor-aware insertion because transcripts append to the end of the buffer.[6] |
| Kimi CLI | None | The repository documents terminal setup, slash commands and MCP tools, and no voice capture.[7] An open issue asks for it; the reply points at macOS Dictation.[8] |
| Pi | Packages | Third-party packages:[9] some run recognition on your machine,[10] one sends audio to a paid service,[11] and send behaviour differs by package.[12] |
A prompt is not a message
A wrong word in a message to a colleague is a typo and they work out what you meant. A wrong word in an instruction gets carried out.
It does not need to know which one you use
There is no integration and no list of supported agents. If it has a text field and it has focus, the words go in. The exception is programs that deliberately reject synthetic input.
Agents people dictate into today — not a compatibility list. The client types into whatever has focus and cannot tell one from another.
What using it actually looks like
There is nothing to install into the agent and nothing for it to support.
Put the cursor where you would type
The prompt line of a terminal agent, a chat box, an editor panel. If a keyboard can write there, so can this.
Press your hotkey and think out loud
Say the whole thing, including the part where you change your mind. The text appears at the cursor about half a second after you stop.
Edit it like anything else you typed
Fix a term, paste a stack trace under it, cut a clause. Still nothing sent.
Send when it says what you meant
Enter, or whatever that agent uses.
Why always-listening modes go wrong
A mode that decides when you have finished has to guess, and users report both failure directions: speech not picked up, and speech cut off mid-sentence.[13] These are user reports on a moving product rather than findings of ours, and any of them may already be fixed.
The underlying reason is worth knowing when you choose a tool: most products do not run their own recognition, they call the browser or operating-system service. That is why voice input in an application can degrade after an update to something that is not the application. A key you press yourself removes the guess entirely — you decide when you have finished, and the recognition is the same one every time regardless of the machine.
What it does not do
- There is no agent integration, no plugin and no MCP server. It types into the focused field, which is why it needs none — and also why it cannot read the agent’s output or act on it.
- It does not read replies back to you. This is dictation, not a voice assistant.
- It does not send. If you want speak-and-go, a built-in voice mode is the right tool and this is the wrong one.
- Transcription happens on our servers, so it needs a connection. If prompts must never leave the machine, run something local instead.
- Dictating prose about code works well. Dictating the code itself does not — punctuation-dense text is the weak case for any dictation.
VibeVoice vs dictating to AI agents: common questions
Does it work with agents that have no voice input?+
Yes, and that is the common case. Nothing is integrated: the client types into whichever text field has focus, so an agent with no voice feature of its own is no different from one that has.
Does it work over SSH?+
Yes. The client runs on your own machine and types into the window in front of you, so it does not matter where the agent is running. Claude Code’s documentation states its own voice dictation requires a local microphone and does not work over SSH or in remote VS Code sessions.
Why not use the voice mode built into the agent?+
Use it if you want the turn to end when you stop talking. The difference is control rather than accuracy: Claude Code’s tap mode submits a transcript of three or more words automatically, while its hold mode can be configured to leave the text in the input. Dictation that only types always leaves it there.
Is there a limit on how long I can speak?+
Not a per-recording one. For comparison, Claude Code’s documentation gives a fifteen-second silence timeout and a two-minute cap for a recording made in tap mode. What you have here is a monthly allowance of minutes, and a four-minute dictation is a normal thing to do.
Does dictating cost extra tokens?+
Not with this. The agent receives text, exactly as if you had typed it, and is billed as text. Speaking directly to a model that accepts audio is a different path with different billing, and you should check that model’s pricing rather than assume.
Is it accurate enough for technical prompts?+
Technical vocabulary usually comes through with no word list to maintain. We publish how we measure accuracy rather than asserting a number against anyone else, and the free tier is there so you can judge it on your own vocabulary.
Sources
Checked 19 August 2026. Every product named here is someone else’s and changes without notice; these are the pages the claims above were read from, so you can see whether they still say it.
- Anthropic — Claude Code documentation, “Voice dictation”code.claude.com
- GitHub Docs — “Use voice input with Copilot CLI”docs.github.com
- Aider — “Voice-to-code with aider”aider.chat
- OpenAI — Codex documentation, “ChatGPT Voice”developers.openai.com
- gemini-cli issue #1982 — “Audio input instead of typing”github.com
- gemini-cli issue #25494 — voice transcriptions append to the end of the input buffergithub.com
- Moonshot AI — kimi-cli repository READMEgithub.com
- kimi-cli issue #1934 — “Feature Request: Voice Input Mode for Terminal and Web UI”github.com
- Pi — package cataloguepi.dev
- Pi — pi-voice-stt package pagepi.dev
- Pi — pi-vox package pagepi.dev
- Pi — picrophone package pagepi.dev
- anthropics/claude-code issue #4542 — voice input truncation reportgithub.com
All of these were read in August 2026. Vendors change what their tools do without changing what this page says, so check the source before relying on any line of the table.
Try it on your next prompt
Whether dictating prompts suits you is a question about how you think, not about the software. 30 minutes a month is enough to find out.
No card. What happens to the audio.
The same client types into VS Code, Slack and Notion. Weighing the tools against each other first? The comparison hub prices seven of them as a customer in Germany pays.
Where to go next
Last updated .