Which MCP servers let Claude and ChatGPT work in WhatsApp?

There are MCP-compatible tools that let AI assistants such as Claude and ChatGPT read, sort, label and draft replies in WhatsApp, and they split into 2 groups: servers that let the assistant send messages on its own, and Kondo Desktop, which stops one step short and leaves sending to you. Which group fits depends on one question: do you want an AI to send WhatsApp messages from your account without you looking at each one?

What is MCP and why does it matter for messaging tools?

MCP (Model Context Protocol) is an open standard for connecting AI applications to outside systems. An MCP server exposes a set of tools with names and typed parameters; an MCP client such as Claude, ChatGPT, Cursor or VS Code calls those tools on your behalf. The protocol’s own documentation describes it as a USB-C port for AI applications: build one server, and every client that speaks MCP can use it.

For messaging, that changes who writes the integration. Before MCP, connecting an assistant to WhatsApp meant a custom integration per assistant. With MCP, the WhatsApp side is written once, and the assistant you already use can list chats, read messages, apply labels and prepare replies through the same tool calls, whether that assistant is Claude Code in a terminal or ChatGPT in a browser.

It also puts the tool list in the open. What an assistant can and can’t do in your WhatsApp is exactly the list of tools the server exposes. If there is no send tool, the assistant can’t send. That makes the tool catalog the first thing to read when comparing options.

Which MCP servers work with WhatsApp today?

The field splits into 2 kinds.

Open-source servers that pair as a linked device and can send. A widely used example is lharries/whatsapp-mcp on GitHub. It runs a Go bridge that links to your personal WhatsApp account through a QR code, stores messages in a local SQLite database, and exposes a Python MCP server with tools to search contacts, list and read chats, and send messages and media. You run it yourself, re-authenticate roughly every 20 days when the linked session expires, and the assistant can send anything it decides to send. The project’s README is candid about the trade-off: an assistant with access to private messages, untrusted incoming content and a way to send is a risky combination if the model is manipulated.

Kondo Desktop’s MCP server, which drafts but never sends. Kondo Desktop is an app for WhatsApp with labels, snippets, reminders and team features, and the Kondo Desktop MCP server exposes 21 tools that cover the inbox: listing and reading chats, searching people, applying labels, archiving, setting reminders, keeping notes on contacts, managing snippets, and saving drafts. There is no send tool, because Kondo wants to protect your WhatsApp account and doesn’t want to be a source of AI spam. The assistant prepares the message in the compose box; you read it and press Send.

If your requirement is “the assistant must send without me,” Kondo Desktop is not the tool, and an open-source server like lharries/whatsapp-mcp is the better fit. If your requirement is “help me get through my WhatsApp inbox without handing an AI the keys to my account,” Kondo Desktop is built for that.

How does Kondo Desktop’s MCP connect WhatsApp to Claude and ChatGPT?

Kondo Desktop runs on macOS, Windows and Linux and pairs with WhatsApp the same way WhatsApp Web does: as a linked device on your phone. Your message history, contact names and attachments stay on your computer and are never uploaded to Kondo’s servers. Only the metadata Kondo creates, such as labels, reminders, snippets and drafts, is stored in your account.

The MCP server keeps that model. Claude or ChatGPT connects to https://mcp.trykondo.com/mcp, signs in with your Kondo account, and each tool call is relayed to the Kondo Desktop app running on your machine, which reads its local database and answers. The relay passes the request and response through in memory and doesn’t keep a copy of chat content. 2 practical consequences follow:

• The Kondo Desktop app has to be open for the tools to work. If your assistant reports no tools, open Kondo.
• Kondo answers from one computer at a time. If you use Kondo Desktop on 2 machines, Settings > MCP shows which one is answering, with a Switch here button.

MCP access is part of Kondo’s Business plan, listed on the pricing page at $36 per seat per month billed annually, alongside CRM integrations and analytics; Enterprise includes it too. Client support covers Claude Code, Claude on the web, the Claude desktop app, ChatGPT, Cursor, VS Code with GitHub Copilot, and any client that can reach a remote MCP server over Streamable HTTP.

The full tool list, grouped the way Kondo’s docs group it:

read_chat returns messages including transcripts of voice notes, so an assistant can summarize one it can’t listen to. set_chat_draft saves a draft that appears in the compose box when you open the chat, and it can also start a first message to a contact you have never messaged. set_chat_reminder archives the chat until the reminder is due, or until the other person replies, then returns it to the inbox.

How do you draft WhatsApp replies with an assistant?

The pattern is: the assistant reads, the assistant drafts, you send. Some prompts that map directly onto Kondo Desktop’s tools:

• “Show me my unread messages.” The assistant calls list_inbox with the unread filter and read_chat on the results, then returns a summary ordered by what needs attention.
• “Draft a reply to the last message in the chat with my accountant. Keep it short and confirm Thursday at 3.” The assistant reads the chat, writes the reply, and saves it with set_chat_draft. You open the chat in Kondo, see the draft in the compose box, edit if needed and send.
• “Draft a follow-up for every chat labeled Lead where I sent the last message more than 3 days ago.” The assistant filters the Lead inbox for chats awaiting a reply, drafts one message per chat, and saves each as a draft. You then go through the drafts one by one.
• “Use my ‘Intro call’ snippet to reply to the 2 new leads.” The assistant reads the snippet with read_snippet, writes in the values it knows from the chat, and drafts. Kondo’s WhatsApp skill tells the assistant to leave anything it can’t fill in the exact {{one to five words}} form, which Kondo catches before sending; other placeholder styles go out as written.

Trust rests on 2 details. First, every draft is visible in Kondo before anything leaves your account, so a bad draft never reaches the other person. Second, in ChatGPT, write actions such as saving a draft or applying a label show a confirmation before they run.

How do you track follow-ups on WhatsApp with an assistant?

Follow-ups are where the labels and reminders do the work, and where an assistant saves the most time because the questions are tedious to answer by hand.

• “Which chats labeled Lead are waiting on my reply?” list_inbox has a needs-my-reply filter, so the assistant lists the chats where the other person sent the last message.
• “Go through my chats from the last 3 days and set reminders for anything I need to follow up on.” The assistant reads recent chats, decides which contain an open promise or question, and calls set_chat_reminder on each. A chat with a reminder leaves the inbox and comes back when the reminder is due, or sooner if the other person replies.
• “Label all recruiter messages as Recruiter.” The assistant reads and classifies, then applies the label with set_chat_label. Labels combine into inboxes, so a “Recruiters” inbox is one set_inbox call after the assistant looks up your labels.
• “Add a note to my accountant’s contact: prefers WhatsApp over email, sends invoices on the 1st.” set_contact_note appends to the note kept on that person, which shows in Kondo next time the chat opens.

Because labels, reminders and snippets live in your Kondo account rather than on one device, they follow you between computers, and snippets can be shared with your team. Labels and snippets are also the same in Kondo for LinkedIn, which has a separate MCP server with its own tool set.

Is it safe to let an AI assistant read your WhatsApp chats?

That question is really 2 questions, and Kondo can only answer the first.

The first is what the tool vendor keeps. Kondo keeps none of your received chat content: message history, contact names and attachments stay on your computer, the relay passes tool calls through in memory, and Kondo’s servers store only the labels, reminders, snippets and IDs Kondo creates. Drafts are the one piece of content Kondo keeps server-side, so a draft an assistant saves is stored in your account until you send or delete it. Nothing an assistant reads through Kondo Desktop is copied to Kondo. The other exception is voice note transcription, which, if you turn it on, sends the audio out for transcription only; it isn’t stored or used for training.

The second is what the assistant keeps, and that is outside Kondo’s control. When Claude or ChatGPT calls read_chat, the messages in that chat go to the model provider and are handled under that provider’s terms, retention and training policies, not Kondo’s. Connecting a hosted model to your WhatsApp opens your inbox to a third party you don’t run, with risk you can’t fully measure from the outside. The same is true of every MCP server in this article, open-source ones included, unless the model itself runs on hardware you control.

So before you connect any assistant to your chats: read the provider’s data policy, check what your workspace admin has agreed to, and decide whether your chats are content you are willing to share with that provider. Kondo can’t make that decision for you, and no vendor’s privacy posture, including Kondo’s, changes what the model on the other end does with what it reads. If the answer is no, a self-hosted model is the only setup where the inbox stays yours end to end.

How do you set up Kondo Desktop’s MCP server?

Requirements: Kondo Desktop installed, signed in and paired with WhatsApp; a Kondo subscription on the Business or Enterprise plan; an MCP client that supports remote servers.

The server address for every client is https://mcp.trykondo.com/mcp. The first time a client connects, you sign in through your browser.

Claude Code

claude mcp add kondo --transport http https://mcp.trykondo.com/mcp

Run /mcp inside Claude Code and follow the prompt to sign in.

Claude on the web

Go to Settings > Integrations, click Add more integrations, then Add custom MCP server, and paste the address.

Claude desktop app

Go to Settings > Connectors, click Add Connector, enter the address and sign in.

ChatGPT

Kondo connects as an app, which requires a ChatGPT Business, Enterprise or Edu plan; Plus and Pro users can connect through Developer Mode but can’t publish the app to a workspace. A workspace Admin or Owner first enables Developer Mode under Settings > Apps > Advanced settings. Then, in ChatGPT, go to Settings > Apps > Create, enter the name Kondo, paste the address and complete the sign-in flow. ChatGPT asks for confirmation before any write action.

Cursor and other JSON-configured clients

{
  "mcpServers": {
    "kondo": {
      "url": "https://mcp.trykondo.com/mcp"
    }
  }
}

Clients that only support local stdio servers can reach the address through the mcp-remote bridge; the Kondo MCP setup guide has the exact command.

After connecting, give your assistant Kondo’s WhatsApp skill so it uses the tools well: what to ask first, how to pace writes, and what it can’t do. The skill is a Markdown file at trykondo.com/skills/kondo-whatsapp.md; save it as a SKILL.md in your assistant’s skills folder or paste it into the assistant’s instructions.

To disconnect a client, go to Settings > MCP in Kondo Desktop, find the client under Connected apps and click Revoke. Access ends immediately.

The Kondo MCP setup guide has screenshots and troubleshooting. Kondo’s source code is private, so there is no public repository to link; the docs page and the skill file are the public reference.

How do you set up an open-source WhatsApp MCP server?

Using lharries/whatsapp-mcp as the example: clone the repository from github.com/lharries/whatsapp-mcp, run the Go bridge and scan the QR code it prints with WhatsApp on your phone, then register the Python server in your client’s MCP config with the path to the checked-out code. Messages sync into a local SQLite database, and the tools become available in Claude or Cursor once the bridge is running. Expect to re-scan the QR code roughly every 20 days, and read the README’s security warning before granting a model send access to your personal number.

Common questions about MCP and WhatsApp

Can Claude send WhatsApp messages for me?

With Kondo Desktop, no. Kondo’s MCP server has no send tool; Claude can read chats, apply labels, set reminders and save drafts, and you send each message yourself. If you need Claude to send on its own, an open-source server such as lharries/whatsapp-mcp exposes a send tool.

Does ChatGPT work with WhatsApp through MCP?

Yes. ChatGPT connects to remote MCP servers as apps on Business, Enterprise and Edu plans, and on Plus and Pro through Developer Mode, once a workspace Admin or Owner has enabled it. Kondo Desktop’s server at https://mcp.trykondo.com/mcp works this way, and ChatGPT shows a confirmation before write actions such as labeling or archiving.

Where do my WhatsApp messages go when an assistant reads them through Kondo?

The assistant’s request goes to Kondo’s relay, which forwards it to the Kondo Desktop app on your computer. The app reads its local database and sends the answer back through the relay to the assistant. The relay passes the exchange through in memory and doesn’t store chat content, and Kondo’s servers never receive your message history.

Does the Kondo Desktop app need to be running for the MCP tools to work?

Yes. The tools are answered by the app on your machine, so if Kondo Desktop is closed the assistant sees no tools or gets no response. Quit and reopen Kondo if commands stop working.

Is Kondo’s MCP included in every plan?

No. MCP is listed under the Business plan on the pricing page, at $36 per seat per month billed annually, with Enterprise including everything in Business. Basic covers labels, snippets and voice notes without MCP.

Can I use one MCP setup for both WhatsApp and LinkedIn?

Not yet. Today Kondo Desktop’s MCP server covers WhatsApp, and Kondo for LinkedIn has its own MCP server with a different tool set and setup guide. One Kondo subscription covers both products, labels and snippets are shared between them, and Kondo is working on one MCP setup for both.

Try Kondo Desktop with your assistant

If you want Claude or ChatGPT to work your WhatsApp inbox without ever sending a message you didn’t read, Kondo Desktop is the MCP server built for that. Try Kondo.