App Logo
Messaging

WhatsApp MCP server for voice notes: download, transcribe and summarize WhatsApp voice messages with Claude or ChatGPT

Prompt template

Find the voice notes contact sent me on WhatsApp this week, transcribe them and summarize what they said.

Get started for free

Overview

Reduck is a universal MCP (Model Context Protocol) server: it turns the websites you are signed in to into tools your agent can call, through browser automation in your own Chrome. For WhatsApp, that is a WhatsApp MCP. To deal with voice notes, your agent lists your chats, reads one, keeps its voice notes and downloads each one as an audio file. Then it transcribes them and summarizes what was said.

Why it's hard

WhatsApp has no API for a personal account. The WhatsApp Business Platform only works for a business number that people write to. WhatsApp's own transcription works one voice note at a time, on your phone, in the languages it supports. The community WhatsApp MCP servers link a new device to your account with a QR code and run a local bridge that you have to install and keep running.

How Reduck does it

Reduck's WhatsApp scripts run in your own Chrome, on the WhatsApp Web session you already have. get_inbox lists your chats without opening them. get_conversation reads a chat's messages, as many as you ask for, where a voice note has the type ptt, without marking anything as read. download_voice_note saves one voice note as an .ogg file in your Downloads folder, checks the file against WhatsApp's own fingerprint, and returns its path, its length and when it was sent. Nothing is played, so a voice note you have not listened to yet stays unheard for the sender. Your agent then transcribes the files and writes the summary.

Setup

  1. Install the Reduck extension in Chrome and sign in. Link WhatsApp Web in that Chrome, then close its tab before you ask your agent: the scripts use the same session, and they stop if a WhatsApp Web tab is open in that browser.

  2. Add the Reduck MCP server, https://mcp.reduck.ai, to your client. It is a remote server over streamable HTTP with OAuth sign-in.

    Claude Code:

    claude mcp add reduck --transport http --scope user https://mcp.reduck.ai

    Claude Desktop and ChatGPT: add a custom connector named reduck with the URL https://mcp.reduck.ai. Other clients: the settings to use.

  3. To transcribe on your Mac, install whisper.cpp and ffmpeg once:

    brew install whisper-cpp ffmpeg

    Then download a model, for example ggml-base.bin from the whisper.cpp models.

  4. Ask: "Find the voice notes Lucile sent me on WhatsApp this week, transcribe them and summarize what she said."

    The first time your agent downloads several voice notes in a row, Chrome can ask whether web.whatsapp.com may download multiple files. Allow it, or the download stops.

Tools

Script What it does Takes
get_inbox Your chats, most recent first, with unread counts and the last message, without opening them a tab (all, unread, groups…), how many
get_conversation A chat's messages, as many as you ask for, with the type of each one (ptt for a voice note) the chat id from get_inbox, how many
download_voice_note One voice note or audio message saved as a file, with its length, format and date, without playing it the message id

Every script only reads. None of them sends a message, plays a voice note or marks a chat as read.

Transcribe and summarize

The script stops at the audio file: the transcript and the summary are your agent's work. In our test, Claude Code converted the .ogg file with ffmpeg, ran whisper.cpp with the base model on the Mac, and got the French sentence right, with the language found automatically. Any speech-to-text tool your agent can run works the same way.

Your agent Find and download voice notes Transcribe
Claude Code, Codex, Cursor Yes Yes, with a tool on your computer (whisper.cpp, Whisper)
Claude chat (claude.ai) Yes No: the chat cannot run a tool on a file of your computer
ChatGPT chat (chatgpt.com) Yes No: the chat cannot run a tool on a file of your computer

WhatsApp MCP servers compared

Community WhatsApp MCP (whatsmeow, Baileys) Kaption AI Reduck
Setup Clone a repo, run a Go or Node bridge, scan a QR code Chrome extension, remote MCP connector One MCP URL, WhatsApp Web linked in Chrome
New linked device Yes Not documented No: it uses your WhatsApp Web session
Downloads voice notes Yes, in most of them Not documented Yes, checked against WhatsApp's fingerprint
Transcription Often built in (Whisper, whisper.cpp) Built in By your agent, on your computer
Sender sees the voice note as played Not documented Not documented No, tested
Runs A bridge that stays connected While WhatsApp Web or its desktop app is open Only when you or your agent asks

Who it's for

  • People who get long voice notes from family, friends or clients and would rather read them
  • Freelancers and small businesses that get voice notes from clients and need them in writing
  • Anyone whose language or accent WhatsApp's own transcription does not handle

FAQ

No. Meta has not published an MCP server for personal WhatsApp accounts. The WhatsApp MCP servers on GitHub and in MCP directories are community projects. Reduck is a universal MCP server, and its WhatsApp scripts give your agent a WhatsApp MCP that runs on your own WhatsApp Web session.
WhatsApp's own transcripts work one voice note at a time, on your phone, in the languages it supports. An agent can take all the voice notes of a chat in one request, transcribe them in any language Whisper knows, summarize them, and keep the text with the rest of your notes.
No. The script downloads the file without playing it and without opening the chat. We tested it on a voice note nobody had played, on 26 October 2026, and the sender still saw the grey microphone. A voice note you already played stays as it was.
The script gives your agent the audio file. To turn it into text, the agent runs a speech-to-text tool on that file, so it needs access to your computer. Claude Code, Codex and Cursor can do it. In the claude.ai or chatgpt.com chat, the agent can find and download your voice notes, but it cannot run Whisper on the file.
The file is saved in the Downloads folder of the computer where your Chrome runs. If your agent transcribes it with whisper.cpp, as in our test, the audio stays on that computer. Reduck itself does not transcribe or keep the audio.
The voice note was sent to you, and with a local tool such as whisper.cpp the audio and the text stay on your computer. If you use it for work, for example with clients' voice notes, tell them that you transcribe them, and check the rules that apply to you.
No. Most WhatsApp MCP servers link a new device to your account with a QR code and run a local bridge built on whatsmeow or Baileys. Reduck uses the WhatsApp Web session you already have in Chrome, so there is no new device and nothing to keep running.
WhatsApp's [Terms of Service](https://www.whatsapp.com/legal/terms-of-service) restrict automated access to its service. Reduck runs only when you or your agent asks, in your own browser, on your own account, and it only reads. We cannot promise that WhatsApp will never flag it. Read WhatsApp's terms and decide whether this fits how you use it.

Similar use cases

WhatsApp

Automate WhatsApp

Read your inbox, reply in context, and send updates or run group polls, no Business API needed.

Messaging
Discord

Discord MCP

A Discord MCP server that acts as your own user account, not as a bot. Reduck gives ChatGPT, Claude or any MCP client tools to read your servers, channels and DMs and reply as you, through browser automation in your own Chrome, with no bot to invite and no user token to paste.

Messaging
Discord

Discord auto reply with your AI agent

Let Claude, ChatGPT or any AI agent automatically read your Discord DMs and the messages that mention you, and answer them from your own account, through the Reduck MCP. Run it on a schedule to reply on its own, or keep a draft-and-approve step. No bot to add, no selfbot, no user token to paste.

Messaging

Start building not clicking

Get started for free