liveface-mcp: Liveface for AI agents

An MCP server for Claude Code, Claude Desktop and Codex. Inside the chat, the studio lets you choose an avatar and a voice, write a line and get a video of the avatar saying it, or make a new avatar from a photo, once you have connected your Liveface account. Without an account, shared avatars still speak in a browser window.

Install

With uv nothing needs installing first. In Claude Code:

claude mcp add liveface -- uvx liveface-mcp

or in the MCP config of Claude Desktop or Codex:

{
  "mcpServers": {
    "liveface": {
      "command": "uvx",
      "args": ["liveface-mcp"],
      "env": { "LIVEFACE_CLIENT_NAME": "Claude Desktop" }
    }
  }
}

Or install it with pipx install liveface-mcp (or pip install liveface-mcp) and use the liveface-mcp command. Python 3.10 or later.

Connecting your account

Ask for the studio and press Connect your account (or, in a terminal, ask the agent to connect_account). A page of the dashboard opens with a code; sign in as usual, check the code is the one the app shows, and press Approve. That is the device sign-in of docs/connected-apps.md: the app never sees your password, gets a session limited to avatars, their creation and speech (never members, keys, deleting, or settings), keeps it in ~/.liveface/mcp-credentials.json (readable by you only), and is cut off at once when you disconnect it in the dashboard's Settings, under Connected apps.

Configuration

The studio (an MCP app)

avatar_studio() returns a page the host draws inside the chat (ui://liveface/studio.html, text/html;profile=mcp-app):

• Library. Your avatars as cards (and the shared ones), a voice (every server voice, grouped by provider), the line ([happy], [sad]... tags change the face), Create video: the page runs the avatar engine, has the avatar say it and records picture and voice into an MP4 (WebM where the browser writes nothing else) to play or download. A browser voice cannot be recorded, so an avatar on one is said in a server voice, and the page says which.
• New avatar. A person (woman, man) or an animal, the photo, the style (the dashboard's own pictures), the pose, the image service, the voice (one that suits the person), a name, and the agreements in the dashboard's own words. The wizard's steps run on this server (creation.py), with progress; the new avatar is then chosen in the library. When its face's points cannot be found (the dashboard has the person place them by hand), the picture is made once more, then Finish in the dashboard opens the wizard where the avatar is waiting.
• The page holds no account token: it asks this server through the host, and the server calls the site. Its CSP lets it reach the site only.
• It needs a host that draws MCP apps. Elsewhere the tools still answer in text, and speak opens its browser window.

Tools

The model's: avatar_studio, connect_account, account_status, disconnect_account, list_avatars, speak(text, avatar), stop. The page's own (hidden from the model): studio_*.

Licence

Proprietary: free to use with your own Liveface account. See the LICENSE file in the package.