Claude Code text to speech: make it read its replies aloud
Claude Code has no text to speech built in. Hear it announce approvals, questions and finished tasks, then compare the ways to make it talk back.
Claude Code can listen: with /voice it takes your prompts by speech. It doesn't talk back. Every reply is still text you have to read, and when Claude waits for an approval it waits silently until you look. If you want to hear what Claude did while you're away from the screen, you add the voice yourself. Here's what exists today, from a one-line hook to a plugin.
Does Claude Code have text to speech?
No. Claude Code's built-in voice feature is voice dictation, which works the other way round: turn it on with /voice, hold Space (or tap) and your speech is transcribed into the prompt. The transcription runs on Anthropic's servers and needs a Claude.ai sign-in. Nothing in Claude Code turns its replies into speech.
The closest built-in option is screen reader mode (claude --ax-screen-reader). It replaces the animated interface with plain, labeled lines that a screen reader such as VoiceOver or NVDA can read. It's made for accessibility: the screen reader reads everything as it appears, code included.
What you actually want to hear
Reading a whole reply aloud sounds better than it is. Replies are full of code blocks, file paths, diffs and tables, and a voice spelling out src/lib/auth/token.ts is noise. What helps when you're away from the screen:
- A short summary when a task finishes: what Claude did, in a sentence or two.
- What it needs when it waits for you: a command to approve, a question to answer, a plan to review. Otherwise it sits there until you come back.
- Several sessions without chaos: voices that queue instead of talking over each other, and that say which session is speaking.
Keep those three in mind for the options below.
Option 1: a Stop hook and your system voice
Claude Code hooks run your own command at points in a session. The Stop hook fires when Claude finishes a reply and receives that reply as last_assistant_message on stdin; Anthropic's hooks documentation names read-aloud hooks as a use for it. On macOS, with jq installed, this in ~/.claude/settings.json reads every reply with the system voice:
{
"hooks": {
"Stop": [
{
"hooks": [
{
"type": "command",
"command": "jq -r '.last_assistant_message' | say",
"async": true
}
]
}
]
}
}
"async": true lets Claude carry on while the voice speaks. On Windows, the built-in Windows voice can do the same job from a PowerShell hook: see Claude Code text to speech on Windows.
It's free, local and set up in a minute. It also reads everything, code included, in the system voice. It speaks only when a reply finishes, so hearing about approvals takes a second hook on the Notification event, and two sessions talk over each other. Our guide to a sound or a spoken summary when Claude Code finishes explains both hook events.
Option 2: a text-to-speech MCP server or skill
MCP servers give Claude extra tools, and several offer text to speech, from ElevenLabs' hosted server to open-source servers that run locally (MCP server, skill or hook? explains the difference). Claude then has a tool it can call to say something. A skill works much the same way: a folder of instructions, often with a script, that Claude loads when it thinks the task calls for it, such as a script that sends text to a speech engine.
The catch is that Claude decides. It speaks when it chooses to, or when your instructions tell it to, and every spoken line is a tool call or a script run inside the turn. It can't announce a permission prompt either: while Claude waits for your approval, it isn't running. Cloud voices also need an API key and send the text to the provider.
Option 3: Verbative Speak, a text-to-speech plugin for Claude Code
Verbative Speak is a Claude Code plugin built for the three things above:
- A short summary when a task finishes. Claude ends each reply with a one-line summary written to be heard, and that line is what you hear: no code, no file paths.
- What Claude needs when it waits for you: an approval for a command or a file, an answer to its question, or a plan to review.
- Several sessions at once: each says which project it's speaking for, and the voices queue.
The voice is natural and runs on your own computer: no API key, and neither your prompts nor Claude's replies go to a speech service. Choose from eleven American and British voices and set the speed. It's quick, too: a one-sentence summary is ready in well under a second. It works in any Claude Code session, the terminal included, on a Mac with Apple silicon or on Windows. Unlike an MCP server or a skill, it runs on Claude Code's hooks, so it speaks at fixed moments, every time, instead of when Claude decides to. Install it in Claude Code:
/plugin marketplace add verbative/verbative-claude-plugins
/plugin install verbative-speak@verbative
The first session downloads the voice once, about half a gigabyte. It's free for 10 spoken summaries a day, and unlimited with Code Advanced.
Option 4: the Verbative extension for VS Code and Cursor
If you work in VS Code or Cursor, the Verbative extension covers both directions. You get the same spoken summaries and approval reminders, and you can talk back: say "listen", speak your prompt (transcribed on your machine, in about 99 languages), and answer permission prompts out loud with "approve", "deny" or "explain". It also runs several Claude Code CLI agents side by side as chats. Our guide to talking to Claude Code walks through that conversation.
Which one to use
| You want… | Use |
|---|---|
| Every reply read aloud, free, set up in a minute | A Stop hook with your system voice |
| Claude to speak only when it decides to | A text-to-speech MCP server or skill |
| A spoken summary and approval reminders in any Claude Code session | Verbative Speak |
| To talk to Claude Code and hear it answer, in VS Code or Cursor | The Verbative extension |
Frequently asked
Does Claude Code have text to speech?
No. Claude Code's /voice is voice dictation: it turns your speech into text for the prompt, and replies stay on screen. To hear them, add a Stop hook that sends each reply to your system voice, a text-to-speech MCP server, or a plugin such as Verbative Speak.
Can Claude Code read its replies out loud?
Yes, with an add-on. A Stop hook can pass each reply to your system voice, and Verbative Speak reads a short summary written to be heard. Claude Code's screen reader mode makes the terminal readable for VoiceOver and NVDA, but a screen reader reads everything, code included.
Is there a text-to-speech plugin for Claude Code?
Yes. Verbative Speak speaks a short summary when a task finishes and tells you when Claude needs an approval, an answer or a plan review, with a natural voice that runs on your computer. Install it with /plugin marketplace add verbative/verbative-claude-plugins, then /plugin install verbative-speak@verbative.
What's the best text to speech for Claude Code?
It depends on what you want to hear. For every reply read in full, a Stop hook with your system voice is free and quick. For a short summary when a task finishes and a reminder whenever Claude waits for an approval, in natural voices that run on your computer, use Verbative Speak. For a cloud voice that Claude calls when it decides to, use a text-to-speech MCP server.
Is there a text-to-speech skill for Claude Code?
You can write one: a skill with a script that sends text to a speech engine. Like an MCP server, it speaks only when Claude decides to use it and stays silent while Claude waits for your approval. Verbative Speak runs on hooks instead, so it speaks every time a task finishes or an approval is waiting.
Does it work in the terminal?
Yes. Claude Code's hooks and plugins work in the terminal, so the Stop hook above and Verbative Speak both speak in a terminal session. The sound plays on the computer where Claude Code runs.
Does text to speech send my code to the cloud?
It depends on the voice. Your system voice in a hook and Verbative's voice run on your computer, so the text stays there. Cloud text-to-speech services, including cloud MCP servers, receive the text they read. Claude Code itself still sends your prompts to Anthropic as usual.
Does it work over SSH?
The sound plays on the machine where Claude Code runs. Over SSH that's the remote server, so you won't hear it there. Text to speech is for sessions on your own computer.