Text to speech for Claude: MCP server, skill or hook?
Three ways to give Claude a voice: a text-to-speech MCP server, a skill, or a Claude Code hook. How each works, where the sound plays, and which to use when.
There are three ways to give Claude a voice, and they behave very differently. An MCP server gives Claude a tool to speak with. A skill gives it instructions and a script. A hook, in Claude Code, runs by itself at fixed moments. Which one you want comes down to one question: should Claude decide when to speak, or should you hear certain things every time?
The short answer
| MCP server | Skill | Hook (Claude Code) | |
|---|---|---|---|
| Who decides when it speaks | Claude | Claude | Claude Code, at fixed moments |
| Speaks while Claude waits for your approval | No | No | Yes |
| Where the sound plays | Local server: your speakers. Hosted: a link or file | Your speakers | Your speakers |
| Best for | Reading something aloud when you ask | A read-aloud routine you reuse | Hearing what's done and what's waiting, every time |
Option 1: a text-to-speech MCP server
An MCP server adds tools to Claude, and a text-to-speech server adds one such as speak or text_to_speech. Claude calls it when it decides to, or when you or your instructions ask it to. You add one to Claude Code with claude mcp add, and to Claude's desktop app under its developer settings or as a connector (Anthropic: MCP in Claude Code).
Where the sound ends up depends on where the server runs:
- A local server runs on your computer, so it can play the audio through your speakers. Its voice can be local too, such as your system's voice or an open model like Kokoro, with no API key and nothing leaving your machine, or a cloud voice with an API key.
- A hosted server runs in the provider's cloud and can't reach your speakers, so it returns the audio as a file or a link. ElevenLabs' hosted server, for example, signs in with OAuth and returns generated speech as a short-lived download link (ElevenLabs). In Claude's apps, custom connectors connect from Anthropic's cloud rather than from your device (Anthropic: custom connectors).
An MCP server is the right fit when you want Claude to read something on purpose: a draft, a summary of a document, the steps of a walkthrough. A hosted one is the way to get audio files in a premium cloud voice.
Option 2: a skill
A skill is a folder with a SKILL.md file: a name, a description, and instructions, often with a script. Claude loads it when the description matches what you're doing, and runs its scripts with Bash (Anthropic: skills). Anthropic doesn't publish a text-to-speech skill, but one is quick to write. In Claude Code, a file at ~/.claude/skills/read-aloud/SKILL.md like this does it:
---
name: read-aloud
description: Reads text aloud with the system voice. Use when the user asks to hear something read out.
---
To read text aloud on macOS, run `say` with the text. On Windows, run
`powershell -NoProfile -Command "(New-Object -ComObject SAPI.SpVoice).Speak('<text>')"`.
In Claude Code the script runs on your computer, so the sound comes out of your speakers. A skill is a good way to make "read this to me" a routine. Like an MCP tool, though, it only speaks when Claude decides to use it.
Option 3: a hook in Claude Code
Hooks are commands that Claude Code itself runs at fixed points in a session, whether or not Claude thinks of it (Anthropic: hooks). Three matter for speech:
- Stop runs when Claude finishes a reply, and receives that reply as
last_assistant_message. Anthropic's documentation names read-aloud hooks as a use for it. - Notification runs when Claude Code wants your attention, such as a permission prompt that's still waiting after a few seconds.
- PermissionRequest runs the moment Claude asks for permission, with the details of what it wants to do.
On a Mac with jq installed, this Stop hook in ~/.claude/settings.json reads every reply aloud:
{
"hooks": {
"Stop": [
{
"hooks": [
{
"type": "command",
"command": "jq -r '.last_assistant_message' | say",
"async": true
}
]
}
]
}
}
On Windows, the same idea works with PowerShell and the Windows voice: see Claude Code text to speech on Windows.
Why a tool can't tell you that Claude is waiting
When Claude asks for permission to run a command or edit a file, it stops and waits for your answer. While it waits, it isn't calling tools, so neither an MCP server nor a skill can say "Claude needs you." Hooks can, because Claude Code runs them, not Claude. That's the difference that matters most when you step away from the screen: a waiting prompt is exactly what you want to hear about, and it's the one moment a tool stays silent.
Beyond a bare hook: Verbative Speak
A Stop hook with your system voice reads every reply in full, code blocks and file paths included, and two sessions talk over each other. Verbative Speak is a Claude Code plugin built on the same hooks, made for listening:
- A one-line summary when a task finishes, written to be heard: no code, no file paths.
- What Claude needs when it waits: an approval to run a command or edit 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.
Its eleven natural voices run on your computer, with no API key, in any Claude Code session, the terminal included. It's free for 10 spoken summaries a day.
Which one to use
| You want… | Use |
|---|---|
| Claude to read a draft or a document aloud when you ask | An MCP server or a skill |
| Audio files in a premium cloud voice | A hosted MCP server |
| Every finished reply in Claude Code, read in full | A Stop hook with your system voice |
| A short summary of each task, and every approval that's waiting | Verbative Speak |
Frequently asked
Is there a text-to-speech MCP server for Claude?
Yes, several. Local servers run on your computer and play speech through your speakers, with a local voice or a cloud voice. Hosted servers, such as ElevenLabs', run in the cloud and return the audio as a file or a link instead of playing it.
What's the difference between an MCP server, a skill and a hook?
An MCP server gives Claude a tool, and a skill gives it instructions and a script; in both cases Claude decides when to use them. A hook is a command Claude Code runs by itself at fixed moments, such as when a reply finishes or a permission prompt is waiting.
Is there a text-to-speech skill for Claude?
Not from Anthropic, but you can write one: a SKILL.md with a description and an instruction to run your system's voice, such as say on macOS. Like an MCP tool, it speaks only when Claude decides to use it.
Can an MCP server tell me when Claude Code needs an approval?
No. While a permission prompt is open, Claude is waiting for your answer rather than calling tools. Claude Code's hooks can, because they run when the prompt appears; Verbative Speak uses them to tell you what Claude wants to run or edit.
Do I need an API key for text to speech in Claude?
Not with a local voice: your system's voice, an open model running on your computer, or Verbative Speak's voices need none. Cloud voices need an API key or a sign-in with the provider.
What's the best text to speech for Claude Code?
For hearing Claude Code while you work, a hook: you hear every finished task and every waiting approval, every time. Verbative Speak does that with a short summary and natural voices on your computer. For having Claude read something on request, an MCP server or a skill.