WACrew

Guide

Voice notes and transcription in a WhatsApp MCP server

How WhatsApp MCP servers handle voice notes: which ones transcribe audio for the agent, which only download the file, and how to send a playable voice note back.

Updated 2 min read
On this page
  1. Three levels of audio support
  2. How WACrew does it
  3. Sending voice notes back
  4. Privacy
  5. FAQ

In many WhatsApp conversations the important part is a four-minute voice note. An agent that cannot hear it misses the request. WhatsApp MCP servers handle audio in three different ways, and it is worth knowing which one you are installing.

Three levels of audio support

LevelWhat the agent getsProjects
Download onlyA file path; the agent needs another tool to transcribelharries/whatsapp-mcp (download_media) and forks
Transcribe on requestText of the voice note, stored with the messageWACrew (transcrever), finereli/whatsapp-mcp (transcribe_audio), BrOrlandi/whatsapp-mcp-local (local, when you turn it on)
Transcribe when readText appears when the agent reads the chatfelipeadeildo/whatsapp-mcp (Workers AI, the first time the agent reads it)

How WACrew does it

WACrew transcribes with Whisper large-v3. On Apple Silicon it runs locally through MLX, loaded on demand and released when idle; elsewhere set TRANSCRIBE_BACKEND=openrouter to use an API. Transcripts are cached in data/<chat>/transcripts/, and that has two effects:

  • buscar searches inside transcripts, so "find where the client mentioned the invoice" works even if they only said it out loud.
  • resumo_do_dia returns every message of a day with audio transcribed, ready to summarize.

For video, ver_video pulls frames with FFmpeg, and assistir_video has a Gemini model watch the video with its audio and return what was shown and said with timestamps. That is useful when a client records their screen while explaining a bug.

Sending voice notes back

WhatsApp plays a voice note only if it is Ogg Opus. lharries/whatsapp-mcp documents this well: send_audio_message converts other formats when FFmpeg is installed, otherwise send_file sends a plain audio attachment. WACrew's responder_midia sends audio as a voice note too, with FFmpeg as a requirement of the project.

Privacy

Local transcription keeps audio on your machine. API transcription sends the audio to the provider you configure. Cloudflare-based servers use Workers AI inside your Cloudflare account. Choose based on what your clients would expect.

FAQ

Can the agent reply with a voice note in my voice? That needs a voice clone, which none of these servers provide. They send audio files you supply.

Does transcription cost money? Local MLX is free on a Mac. API transcription is billed by the provider; felipeadeildo's README estimates about $0.26 per thousand voice notes on Workers AI.

Which languages work? Whisper handles many languages, including Portuguese, Spanish and English, and detects the language per message.