text-to-speech skill (ElevenLabs) for Claude: voiceovers from text

• Python, JavaScript and CLI examples that write speech to an mp3 file

  The agent writes a short script with the ElevenLabs client, picks a voice ID and model, and saves the returned audio to output.mp3. The CLI example elevenlabs say plays text at once with the default voice and the eleven_v3 model.

  Try: “Turn this paragraph into an mp3 with a calm male voice.”

  Runs code · Quick start code

• Six models trade quality, language count and speed

  eleven_v4 is the default. eleven_v4_turbo, eleven_flash_v2_5 and eleven_flash_v2 are for low latency. eleven_multilingual_v2 and eleven_v3 are older models.

  Models

• Four named voices, or list all voices from the account

  The skill names George, Sarah, Daniel and Charlotte with their voice IDs, and shows how to list every voice on your account.

  Network · Voice IDs

• Stability, similarity boost, style, speed and speaker boost

  Sliders for how steady or expressive the voice is. Eleven v4 supports only stability and similarity boost, so pacing and emotion come from audio tags such as [whispers] and from punctuation. voice-settings.md adds ranges, defaults and presets for narration, chatbots, news and character voices.

  Voice settings

• Pass a language code to guide pronunciation

  The agent can set an ISO 639-1 code such as fr. Unsupported codes are ignored, and eleven_multilingual_v2 does not support the option.

  Language selection

• Controls how numbers and dates are spoken

  auto, on or off decides whether 01/15/2026 is read as a date or character by character.

  Text normalization

• Filter, distance, reverb, background noise and panning

  Adds effects such as a phone filter or a hall reverb. The skill says effects do not change request cost, and that reverb or panning needs a stereo-capable output format.

  Network · Audio effects

• Keeps long audio smooth across several requests

  Passes the text before and after each chunk so the joins have no pops or tone shifts.

  Network · Request stitching

• 13 formats from mp3 to raw PCM and telephony codecs

  Lists the format names for web audio, real-time processing and phone systems, and which plan tier some of them need.

  Output formats

• Receive audio as it is generated, over HTTP or WebSocket

  SKILL.md shows the stream method. streaming.md adds two WebSocket routes: the Text to Dialogue WebSocket for eleven_v4_turbo (about 100ms, one voice per connection) and the stream-input WebSocket for the Flash models and eleven_multilingual_v2, which does not take eleven_v3 or eleven_v4. Its playback example pipes audio into ffplay, and both WebSocket examples send your API key to ElevenLabs.

  Runs code · Streaming

• Common API errors and how to read character usage

  Covers 401, 422 and 429 errors and reading the character count from response headers.

  Network · Error handling and costs

Watch out for

• Uses your ElevenLabs key and sends your text out

  Scripts the agent writes call the ElevenLabs API with your key, so the text you give it is sent to ElevenLabs. Use a key you can revoke. github.com

• Agent may install packages and the CLI

  SKILL.md and installation.md tell the agent to run npx -y @elevenlabs/cli@latest onboard init (unpinned, latest version), and the guide lists npm, Homebrew, Scoop and a curl shell installer. streaming.md's playback example starts ffplay. Read each command before you approve it. github.com

• The key may end up in your project

  SKILL.md says onboard init stores a key for the project, and the README says its onboarding skill puts one in your project's env file. Keep that file out of git. installation.md says elevenlabs auth login stores credentials in the OS keyring instead. github.com

• Requests use your character quota

  The skill has a Tracking Costs section that reads the character count from response headers. Check your ElevenLabs plan limits before long jobs. github.com