Miragea MCP: voiceover, voice cloning and captions for Claude, Cursor and Codex

An MCP server that gives your AI agent a real voice toolkit: text of any length to speech in your saved or cloned voice, exact durations for every file, readable SRT/VTT captions from any local audio or video, and per-scene narration with durationInFrames and word-level captions for Remotion and other video timelines. Local stdio server, one line of config, no GPU. Voice and speech recognition by Miragea.

{
  "mcpServers": {
    "miragea": {
      "command": "uvx",
      "args": ["miragea-mcp"],
      "env": { "MIRAGEA_API_KEY": "your_key", "MIRAGEA_VOICE_ID": "optional_default_voice" }
    }
  }
}

Claude Code: claude mcp add miragea -e MIRAGEA_API_KEY=your_key -- uvx miragea-mcp. No uv? pip install miragea-mcp and use "command": "miragea-mcp". Check everything with uvx miragea-mcp --check.

Then just ask:

• "Read this changelog aloud in my voice and save it as release-notes.mp3."
• "Make captions for ~/Videos/talk.mp4." / "TikTok-style captions for clip.mov."
• "Narrate these 5 scenes for my Remotion video and set each scene's durationInFrames."

Tools

What it does for you automatically

• Long text: chunked at sentence boundaries and stitched with a natural 120 ms pause. You never see the 2,000-character limit.
• Durations: measured from the audio itself, so video timelines line up to the frame.
• Captions: speech-recognition segments are rebuilt into whole words and re-flowed with broadcast rules.
• Uploads: video files are reduced to a 16 kHz mono track before upload (fast, small).
• Caching: the same text, voice and model is never billed twice; the same file is never transcribed twice.
• Safety: existing files are never overwritten (name-2.mp3), the key is only read from the environment and never logged, failed audio tasks are refunded by Miragea, and network retries reuse the same idempotency key so a retry cannot double-charge.
• Clear errors: every error comes back as {"error", "how_to_fix"} for the agent to act on.

Compared with other voice MCP servers

Sources: ElevenLabs MCP README (deprecation notice) and the remotion-superpowers changelog (base-folder restriction, no duration returned).

Configuration

Requirements: Python 3.10+. ffmpeg is optional (without it audio is saved as WAV and videos are uploaded whole).

FAQ

Which MCP clients work? Any stdio MCP client: Claude Desktop, Claude Code, Cursor, Windsurf, VS Code (Copilot agent mode), Codex CLI, Gemini CLI, Cline, Zed.

How do I use my own voice? Save your voice once in your Miragea account and set MIRAGEA_VOICE_ID, or pass reference_audio_url (a public link to a clean 5-20 s clip) with reference_text (its exact words, which uses the cheaper 1x rate). A direct upload of local reference clips is planned by Miragea; until then the reference must be a public URL.

What does it cost? Speech is billed per character, captions at 3 points per audio second (live rates: GET https://miragea.ai/v1/capabilities). Cached repeats are free.

Is my key safe? It is read from the server's environment only, never written to disk, logged or returned.

Can I clone anyone's voice? No. Only clone your own voice or one you have permission to use.

Where are files saved? Where you ask, or in MIRAGEA_OUTPUT_DIR (default ~/Miragea). The cache lives in ~/Miragea/.cache and can be deleted at any time.

Related

• miragea-skills: Agent Skills for screen-recording voiceover, Shorts, captions and Remotion narration.
• Retake: fix a flubbed voiceover line in your own voice (CLI + MCP).
• HeadVox: script to voiceover with word-aligned captions.

MIT License.