Ask Claude about your customer conversations: connecting an AI client to your inbox with MCP

The MonoChat MCP Server lets an AI client such as Claude Code or Cursor read your customer conversations through the Model Context Protocol. You create an access key in your profile, add it to the AI client, and then ask in plain language: “summarize what customers asked about delivery this week”, “find the conversations where someone mentioned a refund”, “which of our WhatsApp templates mention the holiday sale?”. The AI client calls six read-only tools and sees only what your own MonoChat user can see.

This guide explains what MCP is in this context, what the server can and cannot do, how to connect Claude Code or Cursor, and the controls that keep it safe.

Two directions of MCP, and why the difference matters

MCP is an open protocol that lets an AI application discover and call tools in other systems. In customer messaging it shows up in two directions:

• Your AI agent calls other systems. The AI agent that answers customers uses MCP tools to look up a CRM record or book a slot. We covered that in MCP for customer messaging.
• Another AI app calls your messaging platform. You, sitting in Claude Code or Cursor, want to work with the conversations themselves. That is what the MonoChat MCP Server is for.

And a third thing that is often confused with both: Meta’s WhatsApp Business Tools MCP helps developers set up the WhatsApp Business Platform. It does not read your team’s conversations.

What you can ask, in practice

Task	What the AI client does
”What did customers complain about this week?”	Lists recent sessions, reads their messages, groups the topics
”Prepare me for the call with this customer”	Searches the customer, reads their last conversations
”Find chats where the AI handed over to a person and summarize why”	Reads sessions and messages, looks for handover patterns
”Draft knowledge base answers from real questions”	Reads messages, extracts repeated questions
”Which templates do we have on this WhatsApp number?”	Lists the message templates of one channel

Because it reads across channels, the same question covers WhatsApp, Instagram, Messenger, Telegram, web chat and SMS conversations in your MonoChat tenants.

The six tools

Tool	What it does
list_tenants	Lists the tenants the connection can access
list_sessions	Lists conversation sessions of a tenant, newest first, optionally only opened, assigned or closed ones
get_session	Returns one conversation session
list_messages	Lists the messages of a session, newest first
search_customers	Searches the customers of a tenant by keyword
list_message_templates	Lists the message templates of one channel

All six only read. A connection has a scope, Read-only (the default) or Read and write; a read-only connection never sees tools that change data.

How to connect Claude Code or Cursor

1. Turn on two-step verification. A connection needs it, by e-mail or an authenticator app, in your profile’s security settings.

2. Create a connection. In your profile, open the MCP connections tab and select Create. Enter a name (“Claude Code on my laptop”), keep the scope Read-only, and choose an expiry: 90 days by default, 365 at most. Enter the 6-digit confirmation code, then copy the access key. It is shown only once and starts with mcp_sk_. You can have up to 10 active connections.

3. Add it to the client. The server is at https://app.monochat.ai/api/mcp-server and expects the key as a bearer token. In Claude Code:

claude mcp add --transport http monochat https://app.monochat.ai/api/mcp-server \
  --header "Authorization: Bearer <access key>"

In Cursor, add a monochat entry with the same URL and header to ~/.cursor/mcp.json. Any client that supports the streamable HTTP transport and custom headers connects the same way.

4. Ask a question. Start narrow: one tenant, one week, one topic. Check a few of the conversations the AI client cites before you rely on its summary.

Limits

Calls are counted per minute: 60 per connection, 300 per account and 600 per tenant, with at most 4 calls at the same time per connection. When a limit is hit, the server answers with status 429 and a Retry-After header, and the client should wait. For a weekly analysis this is plenty; to keep a full copy of one conversation, use the conversation export (JSON, Markdown, CSV or Excel) instead.

Security: what keeps it under control

Giving an AI client access to customer conversations is a data protection decision, not only a technical one. The controls:

• It acts as you. A connection reaches the tenants your user can access and nothing else, so it can only see what your role lets you see. An agent’s AI client sees what that agent sees, not the whole company. See team inbox roles and permissions.
• Keys are personal, scoped and expiring. Each key belongs to one user, has a scope and an expiry, and can be revoked with Revoke in the MCP connections tab. A revoked or expired key cannot be used again.
• Admins can switch it off. Under Settings › Tenant › MCP Server, a user who can change the tenant’s settings can Turn off MCP access for this tenant: every MCP call to it is then refused, from every connection of every user.
• Every call is logged. The same tab lists the MCP calls made to the tenant, newest first, failed ones included: the user, the connection, the tool, the status, the IP address and the time. This log arrived in MonoChat 3.8.39 (October 2026).
• Customer text is data, not instructions. Messages come from customers and agents. Claude Code’s own documentation warns that servers which fetch external content can expose you to prompt injection, so treat the AI client’s actions with the same care as any tool that reads text written by others, and keep the scope read-only.

Before you connect, check what your privacy notice and your contracts say about processing customer messages with AI services, under the law that applies to you (GDPR, UAE PDPL, India’s DPDP Act and others). Our guide to AI chatbot data privacy lists the questions to ask.

Where to start

Create a free account at app.monochat.ai, connect your channels, and add the server to the AI client you already use. The MCP integration page covers the other direction, connecting MCP tools to your own AI agents, and the developers page lists the APIs.

Sources

• MonoChat help center: Connect AI apps (MCP) (tools, scopes, limits, tenant controls)
• MonoChat help center: Profile settings, MCP connections
• Claude Code documentation: Connect Claude Code to tools via MCP (claude mcp add --transport http with --header, prompt injection warning; opened October 31, 2026)
• MonoChat changelog 3.8.37 and 3.8.39