zvex MCP server

AI video dubbing as an MCP tool: hand it a video URL, get back a fully dubbed video in Russian, English or Spanish — keeping the original speakers' voices through per-speaker voice cloning.

Powered by zvex (声桥).

The server is a thin stdio client over zvex's HTTP API, so it runs anywhere Python does — no GPU, no local models.

Install

uvx zvex                # run it without installing anything
# or
uv tool install zvex    # install the `zvex` command
# or
pip install zvex

(Working from a source checkout instead? Use uv tool install ./mcp/zvex.)

Get an API key

1. Sign in at https://tts.xalhar.top
2. Open Account → API keys and create one
3. Copy the zvex-… value — it is shown only once

Jobs are billed from the same credit balance as the web app, at a flat 10 credits per minute of video. A failed job is refunded in full.

Configure your MCP client

Claude Desktop (claude_desktop_config.json) or Cursor (.cursor/mcp.json):

{
  "mcpServers": {
    "zvex": {
      "command": "uvx",
      "args": ["zvex"],
      "env": {
        "ZVEX_API_KEY": "zvex-your-key-here"
      }
    }
  }
}

Tools

Typical flow:

submit_dubbing_job("https://example.com/episode-01.mp4", target_language="ru")
  → {"job_id": 42, "credits_cost": 240, "duration_sec": 1441.0, …}
wait_for_job(42)
  → {"status": "completed", "final_video_url": "…", "subtitle_url": "…"}

Notes

• video_url must be a publicly reachable http(s) link, up to 500 MB.
• target_language depends on the deployment (ru, en, es on the hosted service).
• tier is fast, standard or professional — it selects which features are available, not the price.
• One job per account runs at a time; a second submission returns HTTP 409.
• Output links are served from the zvex domain and require being signed in there.

Development

uv venv --python 3.11 .venv-test
uv pip install --python .venv-test/bin/python -e .
ZVEX_API_KEY=zvex-… .venv-test/bin/python -m zvex.server