mcp.film

mcp.film / Voice & Dialogue / Cartesia MCP

Cartesia MCP

OfficialRemoteby Cartesiafreemiumverified Sep 22, 2026

Official Cartesia (Sonic) server: low-latency TTS, cloning, voice change, localization.

What it does

Cartesia's official MCP server exposes its Sonic voice stack: TTS with speed, volume and emotion controls, speech-to-text, voice cloning, voice changing, voice localization into other languages, and pronunciation dictionaries. For a filmmaker-agent it is strongest on ADR-style voice replacement and on holding a character's pronunciation of invented names consistent across every line and language version.

Connect

Remote endpoint (Streamable HTTP)
https://mcp.cartesia.ai/mcp
Claude Code
claude mcp add --transport http --scope user cartesia-mcp https://mcp.cartesia.ai/mcp
Claude Desktop (claude_desktop_config.json)
{
  "mcpServers": {
    "cartesia": {
      "command": "npx",
      "args": [
        "-y",
        "mcp-remote",
        "https://mcp.cartesia.ai/mcp"
      ]
    }
  }
}
Cursor (.cursor/mcp.json)
{
  "mcpServers": {
    "cartesia": {
      "url": "https://mcp.cartesia.ai/mcp"
    }
  }
}
Auth: oauth · env CARTESIA_API_KEY — Hosted endpoint is browser OAuth and mints an org key for you; the local uvx route takes CARTESIA_API_KEY from https://play.cartesia.ai/keys

Sample tools

text_to_speechspeech_to_textlist_voicesget_voiceclone_voiceupdate_voicevoice_changelocalize_voiceadd_voice_accentsdelete_voice_accentcreate_pronunciation_dictdownload_fileget_credit_usage

Field notes

The August caveat is gone: OAuth discovery now works. /.well-known/oauth-authorization-server returns 200 on 22 Sep 2026 with an authorization-code + PKCE (S256) flow, dynamic client registration and an `mcp` scope, so standards-compliant clients complete the handshake; /.well-known/oauth-protected-resource is still a 404, which is the older half of the spec and matters only to clients that insist on it. The endpoint itself answers 401 unauthenticated. Development restarted hard — PyPI cartesia-mcp 0.22.3 shipped 21 Sep 2026 (was 0.19.0 in August), with voice accent attach/detach tools in 0.21.0, rate limiting and session capping on the hosted transport in 0.20.0–0.22.2, and a python-multipart CVE bump on 9 Sep. The README now recommends hosted over local. Tool count is 18, counted in cartesia_mcp/server.py, not the 20 the README claims; still no audio `infill`, so do not plan an ADR patch workflow around it. Local route: `claude mcp add -e CARTESIA_API_KEY=<your-api-key> cartesia-mcp -- uvx cartesia-mcp`, needing Python 3.13+, which uvx will fetch and which breaks older pinned toolchains. Repo still carries no LICENSE file and still only ~14 stars despite being first-party. get_credit_usage needs a separate CARTESIA_ADMIN_API_KEY; admin and standard keys are not interchangeable. TTS defaults to save=true, returning a file_id and a download_url that expires in 24 hours.

Also in Voice & Dialogue

Used in

Character-Consistent Series · Build a character-consistent series · Create avatar, talking-head, or UGC-style video · Generate voice, score, and sound assets

Pairs well with