WACrew

Guide

How to integrate WhatsApp with ChatGPT, and the catch with MCP

How to integrate WhatsApp with ChatGPT without a bot service. ChatGPT only connects to remote MCP servers, so here are the two routes that work with a self-hosted WhatsApp history: Codex, or a custom GPT on the WACrew API.

Updated 3 min read
On this page
  1. The catch: ChatGPT only reaches remote servers
  2. Route 1: Codex, OpenAI's coding agent
  3. Route 2: a custom GPT on the WACrew API
  4. Which one to pick
  5. Questions

Most guides on how to integrate WhatsApp with ChatGPT end up selling a hosted bot. If what you want is ChatGPT reading your real client conversations and helping you answer them, there is one detail to get right first: how ChatGPT talks to outside tools.

Checked on November 16, 2026.

The catch: ChatGPT only reaches remote servers

ChatGPT's developer mode gives it full MCP client support, but you add servers by URL: it connects to remote MCP servers over HTTPS. It does not start a local process on your computer. Developer mode is available on paid plans (Plus, Pro, Business, Enterprise, Edu), and work accounts may need an admin to turn it on.

WACrew's MCP server (wacrew-mcp) runs locally next to your agent, the way Claude Code, Codex and Cursor expect. So ChatGPT cannot load it directly. Two routes do work.

Route 1: Codex, OpenAI's coding agent

Codex accepts local MCP servers in its config.toml, so it gets all 37 WACrew tools: read groups, transcribe voice notes, see screenshots, draft and send replies after your approval.

[mcp_servers.wacrew]
command = "npx"
args = ["-y", "wacrew-mcp"]
env = { WAC_CLOUD_URL = "https://your-panel.example.com", WAC_CLOUD_USER = "admin", WAC_CLOUD_PASS = "change-me" }

Full setup in WhatsApp MCP for Codex.

Route 2: a custom GPT on the WACrew API

The WACrew panel has a documented API v1 with Bearer tokens scoped read or write, and it publishes its own OpenAPI spec at /api/v1/openapi. A custom GPT Action can import that spec and call your panel.

  1. Put the panel behind HTTPS (Railway gives you a domain).
  2. On the server, create a token: node scripts/api-token.mjs criar chatgpt --escopos read.
  3. In the GPT editor, add an Action, import the OpenAPI from your panel and set authentication to API key, Bearer.
  4. Ask: "Which chats are waiting for a reply?" (/radar) or "Show the last 50 messages of the Acme group".

Start with a read-only token. The API does not enforce the per-chat confirm mode: it returns the chat mode, but a write token sends right away.

Which one to pick

You wantUse
ChatGPT to triage and summarizecustom GPT with a read token
An agent that drafts and sends in your voiceCodex (or Claude Code, Cursor) with the MCP
No coding agent at allthe WACrew panel itself, with its built-in copilot

Questions

Can I paste the WACrew MCP into ChatGPT's connector screen? Not as is: that screen wants a public HTTPS MCP URL, and WACrew's MCP server is a local process.

Will ChatGPT answer my clients on its own? Not with these routes. The custom GPT answers you, inside ChatGPT. Replying to clients from your number is the job of the agent with the MCP, and each chat stays in confirm mode until you choose otherwise.

Does the custom GPT see voice notes? It sees the transcription once WACrew has transcribed the audio. Media files are served by /media/{path} with the same token.

Is it safe to expose the panel? Keep the panel behind login and HTTPS, use scoped tokens and revoke them when done. See WhatsApp MCP security.