Paxa Labs

What Paxa Labs MCP server does

The Paxa Labs MCP server exposes speech and document-processing functions from the Paxa Labs API to MCP-compatible agents. Its language features cover Thai, English, and Mandarin Chinese speech, while translation can accept any source language and produce Thai output. The server also works with local PDFs and common image formats for OCR and typed-field extraction.

Speech can be handled interactively or as files. Short text can be synthesized and played immediately, while longer content is split into segments and managed through a shared playback queue. Audio recordings can be transcribed in Thai, English, or mixed-language content, with optional speaker identification, word timing, and subtitle output.

How it works

The server is launched with the npm package through npx -y @paxalabs/mcp. MCP clients send tool calls to the local process, which uses PAXA_API_KEY to access the Paxa Labs API. Local audio and document paths are supplied to the relevant tools; generated audio, OCR results, extracted fields, and transcription outputs are returned or written as specified by the tool.

All generated playback uses one ordered audio queue. Short speak calls can be inserted ahead of longer queued material, while queue_speech synthesizes upcoming segments as the current segment plays. Playback controls can report queue status, pause or resume audio, skip an item, or clear the queue. A streaming-capable player can begin playback before the complete audio response has downloaded.

Setup and configuration

Create a Paxa Labs account and obtain an API key, then add the server to the MCP client configuration. A typical stdio entry is:

{
  "mcpServers": {
    "paxa": {
      "command": "npx",
      "args": ["-y", "@paxalabs/mcp"],
      "env": { "PAXA_API_KEY": "pxa_your_key_here" }
    }
  }
}

Claude Code can add the same server with claude mcp add paxa -e PAXA_API_KEY=pxa_your_key_here -- npx -y @paxalabs/mcp. The README also provides installation links for Cursor and VS Code. The package requires Node.js, and the API key is needed for tools that call the service. The repository describes the tool set as beta, so names and behavior may change before version 1.0.

For Claude Code voice notifications, the package can also be installed globally so the paxa say command is available to hooks. Optional playback programs such as mpg123 or ffmpeg can reduce startup delay for cached notification phrases.

Tools and capabilities

The Paxa Labs MCP server includes:

• speak, queue_speech, control_playback, and play_audio for local audio playback.
• text_to_speech for creating MP3, Opus, or WAV files without playing them.
• translate_to_thai with formality, glossary, and context options.
• ocr_document for Markdown or structured-block output from PDFs and PNG, JPEG, or WebP images.
• extract_fields for typed values such as Thai IDs, dates, amounts, banks, and line items; unavailable values are returned as null with a reason.
• transcribe_audio for Thai, English, or mixed recordings, including optional speakers, word timings, SRT, and VTT subtitles.
• list_voices, list_models, and get_account for available voices, model details, pricing, credit balance, plan, and rate limits.

Credit charges vary by tool. Playback control and informational tools are free; speech, translation, OCR, extraction, and transcription use the rates listed by the service.