Paxa MCP: voice and OCR for agents
Paxa MCP
The agent speaks through your speakers, listens on the mic, translates across 15 languages and reads Thai PDFs via Paxa Labs
- 10 ★ on the repository
- Last commit
- MIT
Extended access
Why this level
- The server records from the microphone and reads local audio, video, PDFs and images, then sends them to the Paxa Labs API
- Claude Code hooks run paxa say on events, and the API key spends credits with every voiceover
Install
- Works in
In your terminal, with SkillFoxx CLI
npx skillfoxx add mcp/paxa-mcpDetects the agents on your machine, shows the access level and pins the version.
Other ways to install
Access and checks
How we reviewExtended accessThe tool runs code or makes network calls and only reads external data.
- The server records from the microphone and reads local audio, video, PDFs and images, then sends them to the Paxa Labs API
- Claude Code hooks run paxa say on events, and the API key spends credits with every voiceover
This is third-party code. Review the repository files before installing.
What it does
The MCP server connects an agent to the Paxa Labs API, a speech service for Thai, English and Chinese. The speak tool voices a short line through the computer's speakers, queue_speech reads long text as a chunked queue, and control_playback pauses, skips and clears it. For files there is text_to_speech, which saves speech as mp3, wav, opus or pcm and can write srt and vtt subtitles. Voices can be described in words. design_voice creates a voice from a description, refine_voice adjusts it, and save_voice stores it under an id. listen records a phrase from the microphone and returns a transcript, and transcribe_audio transcribes recordings and video with speaker labels and a list of pauses for editing. The rest deals with documents and text. translate converts between 15 languages with a glossary and formality, ocr_document reads PDFs and images into Markdown, and extract_fields pulls typed fields (Thai IDs, dates, amounts, invoice lines). A separate paxa say command voices Claude Code notifications through hooks.
Who it is for. Video and podcast creators, Thai-speaking teams and anyone who would rather hear an agent's reports than read them and dictate answers.
Good fit when
- The agent finished a long task, you stepped away from the screen and want the result in two spoken sentences
- A lecture video has to become text with srt subtitles and a list of pauses worth cutting
- A voiceover script goes into one audio file with time-aligned subtitles
- Scanned Thai documents are read through OCR and split into fields
- Claude Code should say aloud that it waits for permission
Not a fit when
- The server runs on a machine without sound or a microphone: speak and listen are useless there
- You need Russian speech: the focus is Thai, English and Chinese, and Russian is among the 15 translation languages but has no voices
- Audio must not leave for an external service: every voiceover and transcript goes through the Paxa Labs API
Example request
Transcribe interview.mp4 with srt subtitles and list the pauses longer than 0.7 seconds that are worth cutting.MCP
- Transport
- stdio
- Authentication
- API key
Environment variables
PAXA_API_KEYrequired, secret- Paxa API key; calls to the service fail without it
PAXA_OUTPUT_DIR- Folder for text_to_speech files, the working directory by default
PAXA_DEFAULT_VOICE- Default voice, foithong by default
PAXA_MIC- Microphone device for listen
Limitations
You need PAXA_API_KEY: new accounts get free credits, after which speech costs 10 credits per 1000 characters and translation 25. The project is in beta, and tool names and behavior may change before 1.0.
listen needs Node.js 22 or newer, a recording utility such as ffmpeg or sox and microphone permission for the app. Video input and joining audio parts need ffmpeg. Streaming playback requires mpg123, ffplay or mpv, and pause does not work on Windows. The paxa say hook fails when the command is not on PATH.
How to disable. Run claude mcp remove paxa or delete the paxa entry from the client config. If you installed the paxa command, remove the Notification and Stop hooks from ~/.claude/settings.json and run npm uninstall -g @paxalabs/mcp.
FAQ
How do I install Paxa MCP in Claude Code?
Run claude mcp add paxa -e PAXA_API_KEY=pxa_your_key_here -- npx -y @paxalabs/mcp. Get the key at paxalabs.com. After that the agent has speak, queue_speech and the other tools.
How do I make Claude speak when it waits for permission?
Install the paxa command with npm install -g @paxalabs/mcp and add a Notification hook with the command paxa say to ~/.claude/settings.json. For a summary at the end of each turn, add a similar Stop hook.
How much do calls cost?
Usage is counted in credits: speech 10 per 1000 characters, translation 25 per 1000 characters, OCR 6.5 per page, transcription 8.33 per minute, microphone listening 12.5 per minute. get_account shows the balance.
Is there an extension for Claude Desktop?
Yes, GitHub releases carry an .mcpb bundle with its own copy of the server. Open it in Claude Desktop and enter the key in the extension settings; Node.js and npm are not needed.
Do I need ffmpeg?
Not for speech. ffmpeg is required for video input and for joining mp3, opus and wav parts, and listen works with either ffmpeg or sox. Without it text_to_speech and the other tools still work.
Discussion
No questions or reviews yet.
Related
MCP serversAn MCP server for Sber's SaluteSpeech voice API: records speech from the microphone, recognizes it and speaks the agent's reply
Local speech synthesis for agents: a command and a skill voice text, clone a voice from a recording and render scores to audio, with no cloud
A text-to-speech MCP server that turns text and multi-voice scripts into audio through ElevenLabs and keeps a job history.
MCP server for Sber SaluteSpeech: transcribe Russian speech to text and voice text with Sber voices
A local MCP server for voice work on macOS: voice cloning, speech synthesis and audiobooks from PDF on an agent's command
Official MiniMax MCP server: text to speech, voice cloning and design, image and video generation
Comparison with similar tools
Cards that share the main search query, and how they differ.
| Tool | Official | Authentication | Repository stars | Main difference |
|---|---|---|---|---|
| Paxa MCP: voice and OCR for agents (this card) | yes | API key | 10 | |
| ElevenLabs Skills | yes | 536 | ElevenLabs voice and speech skills, while this is a Paxa server with OCR, translation and a playback queue | |
| SaluteSpeech MCP | no | API key | 1 | SaluteSpeech recognition and synthesis for Russian, while this focuses on Thai, English and Chinese |
README in short
The description gives a quick start for Claude Code, Cursor, VS Code and Claude Desktop, a table of 19 tools with prices in credits, a section on voice mode and hooks for Claude Code, a description of the Paxa TTS Flash v2 and v1 speech models with direction tags like [whispering], and tables of environment variables and players per platform. The license is MIT.