@aiwerk/mcp-server-elevenlabs

An MCP server for the ElevenLabs API. 390 tools, generated from the live OpenAPI document, covering every documented endpoint.

Why this exists

ElevenLabs ships its own MCP server. It is hosted, it authenticates with OAuth, and it exposes about a dozen high-level tools for managing ElevenAgents plus one text to speech call. That is a good fit for "create a support agent and set its voice".

This one is for the rest of the API. The whole of it:

It also handles the two things a generated client usually gets wrong: 31 endpoints take file uploads (speech to text, voice cloning, dubbing, audio isolation, knowledge base) and 21 return raw bytes (all the text to speech variants). Parsing audio as JSON produces a plausible-looking string of mojibake, so those paths are written by hand and tested.

What the official hosted server does better: OAuth means no API key is copied into the client, and its agent tools are composed product logic rather than raw endpoints ("what would this agent cost per conversation on a different model" is not one API call). Use both if that is what you need. They do not conflict.

Related MCP server: ElevenLabs MCP Server

Install

npm install -g @aiwerk/mcp-server-elevenlabs

Or run it straight from npx in a client config:

{
  "mcpServers": {
    "elevenlabs": {
      "command": "npx",
      "args": ["-y", "@aiwerk/mcp-server-elevenlabs"],
      "env": {
        "ELEVENLABS_API_KEY": "your-key",
        "ELEVENLABS_OUTPUT_DIR": "/where/audio/should/land"
      }
    }
  }
}

Get a key at https://elevenlabs.io/app/settings/api-keys. The free tier includes 10k credits a month.

Configuration

Files in and out

Uploads. Every binary field is offered two ways:

// local install: the server can read your disk
{ "file_path": "/home/me/interview.mp3", "model_id": "scribe_v1" }

// containerised or remote: send the bytes
{ "file_base64": "SUQzB...", "file_filename": "interview.mp3", "model_id": "scribe_v1" }

Fields that accept several files (add_voice, create_finetune, add_pvc_voice_samples) use files_paths / files_base64_list / files_filenames.

Downloads. Anything returning audio, video or a zip takes output_path:

{ "voice_id": "...", "text": "Guten Tag", "output_path": "greeting.mp3" }
// → { "contentType": "audio/mpeg", "bytes": 26375, "path": "/output/dir/greeting.mp3" }

A relative path resolves against ELEVENLABS_OUTPUT_DIR. With no path and no output dir, the audio comes back as an MCP audio block, as long as it is under the inline limit. Base64 inflates by a third and every byte crosses the model's context, so prefer a file for anything longer than a sentence.

Credits and safety

66 operations spend credits, and each one says so in its description. Nothing here guesses on your behalf:

• ELEVENLABS_DRY_RUN=1 blocks every write and generation call.

• Only GET is marked read-only. Several POSTs merely query, but every one of them also bills, so they are gated with the writes.

• DELETE operations carry destructiveHint.

An ElevenLabs key can be restricted per endpoint group, given its own credit quota and locked to an IP range. All three failures arrive as HTTP 401, and this server tells them apart, so "out of credits", "this key may not touch this endpoint" and "this host is not on the allowlist" do not all read as "check your credentials".

Regenerating from the spec

The spec is committed, not fetched at build time, so a vendor edit cannot land in a release nobody reviewed.

npm run fetch-spec   # writes spec/elevenlabs-openapi.json, reports whether it changed
npm run regen        # naming table + generated tools
npm test

Development

npm install
npm run build
npm test             # 47 unit tests
npm run smoke        # live test against the real API, spends a few credits

The smoke test does a round trip that the unit tests cannot: it generates speech to a file, then uploads that same file back to speech-to-text and checks the words come out again.

License

MIT. Not affiliated with ElevenLabs.