Compare · Claude Code text-to-speech · 8 min read
How to make Claude Code talk: every text-to-speech option compared.
Claude Code’s /voice lets you talk to it. In the terminal, nothing built in talks back. Here are the ways people add spoken output, what each one actually reads, and which to pick.
say hook · claude-code-narrator · claude-code-tts · VoiceMode · Aloud · Heard · Checked October 8, 2026
Updated
The short answer
Text-to-speech for Claude Code: add a say hook, a plugin, VoiceMode, Aloud or Heard; /voice is input only.
Claude Code’s /voice lets you talk to it; Heard (heard.dev), the voice context layer for agents, makes it talk to you: it speaks questions, approvals, results and failures from every Claude Code and Codex session on macOS, and on Heard Power you talk back. The terminal CLI has no spoken output of its own, since voice dictation only turns your speech into the prompt. Other add-ons are a Stop hook that pipes Claude’s last message into macOS say; a read-aloud plugin such as claude-code-narrator, ybouhjira’s claude-code-tts, or the Heard team’s free heardlabs/claude-code-tts (a Stop hook and a /say skill using your computer’s built-in voice; a separate repo, not the Heard app); VoiceMode, for a two-way spoken conversation; or a Mac tool that speaks alerts and short gists instead of whole replies, such as Aloud.
The difference that matters is what gets read. Most hooks and read-aloud plugins speak Claude’s own words, which is accurate but long (ybouhjira’s claude-code-tts reads just the first sentence). Aloud (free, local Kokoro, switched on per session) and Heard put questions, approvals and failures first and shrink the rest to a sentence or two. Heard’s Mac app sets up the hook for you; the open-source engine is free, and the app has paid plans after 14 days of free Heard with no card. Where the others win: cost and platforms. None of the other options charges anything itself; ybouhjira’s claude-code-tts needs your own OpenAI API key, which OpenAI bills, and the rest can all speak with a local voice at no cost. Heard runs only on macOS 14 or later; the Platforms column shows which of the others also run on Linux or Windows.
Outside the terminal, users report that the Claude desktop app’s Code tab added a play button in 2026 that reads one response on demand; see #58429.
All options
Seven ways to hear Claude Code.
From each project’s README. Stars and dates move; Checked October 8, 2026.
| Option | What you hear | Voice engine | Platforms | Price and license |
|---|---|---|---|---|
| say hookStop hook + jq | Claude’s final message, word for word | macOS say | macOS (Linux: spd-say) | Free |
| claude-code-narratorplugin | Tool steps, text between them, the start of the final reply | Kokoro, local | macOS, Linux | Free, MIT |
| claude-code-ttsOmnicapable | Every reply, skipping code blocks and tables | Kokoro ONNX, local | macOS 12+, Windows | Free, MIT |
| claude-code-ttsybouhjira · MCP | The first sentence of every reply, plus a speak tool Claude can call | OpenAI tts-1, needs an API key | macOS, Linux, Windows | Free, MIT; OpenAI bills about $0.015 per 1,000 characters |
| claude-code-ttsheardlabs · ours | Each reply with Markdown stripped, up to 600 characters, or a short /say summary | Your computer’s built-in voice | macOS, Linux, Windows | Free, MIT |
| VoiceModeMCP + plugin | A spoken conversation you can answer | Local Whisper.cpp and Kokoro, or OpenAI | macOS, Linux, Windows | Free, MIT |
| Aloudsoftcane | Questions, approvals and failures, plus a short gist of each reply; the full reply on request | Kokoro, local | macOS | Free, MIT |
| Heardvoice layer | A sentence or two: what changed and what needs you | Kokoro local, your own keys, or managed voices | macOS 14+ | Free engine, Apache 2.0; paid app plans |
The two-minute version: pipe the last message into say.
The Stop hook receives JSON on stdin that includes last_assistant_message. Extract it with jq and hand it to say. "async": true keeps Claude Code from waiting on the speech, and head -c stops a long reply from talking for minutes.
It costs nothing and needs no install beyond jq. The limits: it reads Markdown, file paths and code exactly as written, which is hard to follow by ear; it speaks only when a turn ends, so a permission prompt needs a separate Notification hook; it never says which project it is speaking for; and with two sessions open, nothing stops their say commands from talking at once. Heard (heard.dev) sets up the hook for you and, with several sessions running, starts each update with the project name.
{
"hooks": {
"Stop": [
{
"hooks": [
{
"type": "command",
"command": "jq -r '.last_assistant_message // empty' | head -c 600 | say",
"async": true
}
]
}
]
}
}
claude-code-narrator and the claude-code-tts repos read replies aloud.
claude-code-narrator is a Claude Code plugin. It speaks short descriptions of tool use (“Reading file settings dot json”), the text between tool calls, the first thousand or so characters of the final reply, and notification text. It goes quiet when you type or click, uses the local Kokoro engine with eight voices, and runs on macOS and Linux.
claude-code-tts by Omnicapable speaks every reply from a Stop hook with local Kokoro ONNX (about a 350 MB download of model and voice files), skipping code blocks, tables, URLs and emoji. It supports macOS 12+ and Windows. A separate npm package of the same name defaults to Edge TTS, which is Microsoft’s online service, with Kokoro as the offline fallback.
A third, claude-code-tts by ybouhjira, is an MCP server with a Stop hook that speaks the first sentence of each reply through OpenAI’s tts-1. It runs on macOS, Linux and Windows, but needs an OpenAI API key, and OpenAI bills about $0.015 per 1,000 characters.
A fourth is ours: heardlabs/claude-code-tts, from the Heard team. A Stop hook reads each reply with Markdown stripped, up to 600 characters, and a /say skill ($say in Codex) speaks a short summary instead of every word. It uses your computer’s built-in voice (say on macOS, spd-say or espeak-ng on Linux, System.Speech on Windows), so there is no account and no API key. It is free and MIT licensed.
/plugin marketplace add shreyas-s-rao/claude-code-narrator
/plugin install narrator
/reload-plugins
/narrator:on
VoiceMode, if you want to talk with Claude, not just listen.
VoiceMode is the most established option here, with about 1,400 GitHub stars and an MIT license. It is an MCP server and Claude Code plugin for back-and-forth voice conversations: Claude speaks, you answer out loud. It can run fully local with Whisper.cpp and Kokoro, or use OpenAI’s cloud. It works on macOS, Linux and Windows.
Pick VoiceMode when the goal is a spoken conversation with one session, or when you are not on a Mac. It is built for talking with an agent, not for keeping track of several agents at once.
claude plugin marketplace add mbailey/voicemode
claude plugin install voicemode@voicemode
/voicemode:install
Aloud speaks alerts and short gists, locally, per session.
Aloud is a free, MIT-licensed macOS tool for Claude Code and Codex. Type aloud on inside a session and it speaks questions, approval requests and failures first, then a short gist of each finished reply. Spoken alerts lead with who is asking, as in its README example “Codex aloud asks: Which database should I use?”. Routine tool calls stay silent, and aloud full reads the whole reply. Speech runs locally with Kokoro. It needs Python 3.11 or 3.12 and Homebrew.
Aloud is the closest option to Heard. It wins if you want speech only in the sessions you choose: each one stays silent until you type aloud on. It installs with pipx and needs no account. Heard speaks for every session of the agents you enable and adds managed voices and, on Power, voice replies.
brew install [email protected] pipx
pipx ensurepath
pipx install --python python3.11 git+https://github.com/softcane/aloud.git
aloud install
Heard summarizes instead of reading, across every session.
Heard installs Claude Code hooks and decides what deserves speech: questions, approvals, failures and results get through; routine tool calls are not read out. With several sessions running, each update starts with the project name. The open-source engine has three listening modes, from short signposts to alerts only; the Mac app has one fixed behavior.
The open-source engine is free under Apache 2.0, with local Kokoro voices and no account. The Mac app adds managed voices and, on Power, voice replies and Parrot recaps. Heard is macOS only.
“Api: tests pass. I fixed the date parser and have one question about the migration.”
“Here is a summary of the changes I made. First, in src slash utils slash date dot ts, I updated the parse function to…”
git clone https://github.com/heardlabs/heard.git && cd heard
python3 -m venv .venv && source .venv/bin/activate && pip install -e .
heard install claude-code
Match the tool to what you want to hear.
All of these are honest choices. They answer different needs.
- Every word of the final reply, free: the
sayhook. - A local narrator on macOS or Linux: claude-code-narrator.
- Replies read aloud on Windows: Omnicapable’s claude-code-tts (local) or ybouhjira’s (OpenAI key, also Linux).
- Free, local alerts and reply gists for the sessions you choose on a Mac: Aloud.
- A spoken back-and-forth conversation, on any OS: VoiceMode.
- A screen reader: Claude Code’s own screen reader mode (
claude --ax-screen-reader), plus any option above if you also want replies spoken without stepping through them with your screen reader. - Short spoken summaries from every Claude Code and Codex session without switching each one on, and voice replies on Power: Heard.
A few more details
Questions?
Is there a text-to-speech option for Claude Code for accessibility?
Yes. If you use a screen reader such as VoiceOver or NVDA, turn on Claude Code’s screen reader mode with claude --ax-screen-reader (or "axScreenReader": true in your user settings): it prints plain, labeled lines instead of boxes and animations, and your screen reader reads them. To have replies spoken without a screen reader, add a Stop hook that pipes Claude’s last message into say, or use a read-aloud tool from this page: claude-code-narrator, a claude-code-tts repo, VoiceMode, Aloud, or Heard (heard.dev), the voice context layer for agents, which speaks a sentence or two about what changed and what needs you on macOS 14 or later. See also Heard for dyslexic builders.
Does Claude Code have text-to-speech built in?
Not in the terminal. /voice is dictation: it transcribes your speech into the prompt, and the audio is processed on Anthropic’s servers. Users report a play button in the Claude desktop app’s Code tab that reads one response on demand. Automatic spoken output in the CLI comes from hooks, plugins, MCP servers or a voice layer.
How do I make Claude Code read its responses out loud?
Add a Stop hook in ~/.claude/settings.json that runs jq -r .last_assistant_message | say, or install a plugin such as claude-code-narrator. Both read Claude’s own words. Heard (heard.dev), the voice context layer for agents on macOS, speaks a short summary of each turn instead.
How do I use the macOS say command with Claude Code hooks, and what are the limits?
Add a Stop hook in ~/.claude/settings.json whose command is jq -r '.last_assistant_message // empty' | head -c 600 | say, with "async": true so Claude Code does not wait for the speech; or let Heard (heard.dev) install its own hook and speak a short summary instead. The limits of say: it reads Markdown, paths and code literally; it speaks only at the end of a turn; it does not name the project; and two sessions can talk over each other.
Is there a free, local text-to-speech option for Claude Code?
Yes. claude-code-narrator, Omnicapable’s claude-code-tts and Aloud all use the local Kokoro engine, our heardlabs/claude-code-tts uses your computer’s built-in voice, VoiceMode can run with local Whisper.cpp and Kokoro, and Heard’s open-source engine uses Kokoro when you add no keys.
Is Kokoro TTS good for reading AI agent updates aloud?
Yes, if you want speech that stays on your machine. Kokoro is an open-weight text-to-speech model with 82 million parameters and Apache-licensed weights, and it runs without a cloud voice service. The model does not decide how much you hear; the tool does, and a long reply takes a long time to hear in any voice. Omnicapable’s claude-code-tts, for macOS or Windows, reads every reply, skipping code blocks and tables; it needs Python 3.9+ and a one-time download of about 350 MB for the Kokoro ONNX model and voice files. Aloud speaks a gist of about 240 characters, with the full reply on request, and keeps Kokoro loaded between events to avoid the startup cost on each message. Heard (heard.dev), the voice context layer for agents, speaks a short summary rather than the full reply, and its Mac app uses managed voices, not Kokoro.
Can I talk back to Claude Code by voice?
Yes: /voice dictates prompts, VoiceMode runs a spoken conversation, and Heard Power lets you hold right ⌘ and name an agent to send that session a spoken reply.
Keep going
Related Heard resources
Give your agents a voice.
Keep the context. Get on with your work.
Includes 14 days of free Heard. No card needed.


