CHBT CHBT chatbot builder
Create a bot
Features
For sellers Shop in your bot AI sales assistant Wildberries Online store QR code for a package insert
Messengers Telegram MAX VK Chat on the site
Templates Plans Blog Help
Sign in Create a bot

Home · Blog · MCP for business: what an AI assistant can do with your chatbot

MCP for business: what an AI assistant can do with your chatbot

· Command CHBT · AI, Integrations

To find out how many leads the bot brought in a week or what yesterday’s customer wrote, you usually have to open the dashboard and click through the sections. With MCP you can ask your AI assistant in plain words: “how many new leads this week?”, “show conversations where the customer is waiting for an operator”, “tell the customer the order has been shipped”. Here is what MCP is, how to connect CHBT to it in five minutes, what the assistant can do and how not to give it too much.

What MCP is in simple words

MCP — Model Context Protocol — is an open standard by which AI assistants connect to external services. A service describes its “tools”: what can be viewed and what can be done. The assistant decides by itself which tool to call to answer your question.

CHBT has its own MCP server. By connecting it to Claude, ChatGPT, Cursor or another MCP-capable client, you talk to the bot through your assistant: it sees your project’s stats, leads and conversations and can reply to a customer.

What the assistant can do through CHBT

Eight tools:

  • get_stats — a summary for 1–90 days (7 by default): conversations, incoming messages, new leads, breakdown by channel. The editor’s test chat is not counted.
  • list_leads — leads: contacts who left a phone or email in the last N days, newest first.
  • list_conversations — conversations filtered by status: run by the bot, run by the AI agent, waiting for an operator, closed.
  • get_messages — the message history of a specific conversation.
  • send_reply — reply to a customer on behalf of an operator: the message goes to the same messenger, and the conversation moves to an operator.
  • handoff_to_operator — hand the conversation to an operator; the bot stops replying in it.
  • close_conversation — close the conversation.
  • list_flows — the bot’s flows: name and whether it is on.

The assistant does not change flows or start broadcasts — there are no such tools.

Step 1. Issue a token

Open API connections in the dashboard, the API tokens section. In the Who do we issue to? field write a clear name, for example “Claude — Anna”, and press Issue a token. The token is shown once — save it somewhere safe. A lost token cannot be recovered: a new one is issued and the old one is revoked.

Tokens can be issued by the bot owner and admin. Make a separate token for each person and each assistant — this way any access can be revoked without touching the rest.

Step 2. Connect the assistant

The MCP server address is shown in the AI assistant: MCP for Claude, ChatGPT and Cursor block on the same page — it is your dashboard address with the path /mcp. The token is passed in the header Authorization: Bearer <token>.

  • Claude Code — a ready-made command in the dashboard: claude mcp add --transport http chbt … with the address and header.
  • Claude, ChatGPT, Cursor and other clients — add an MCP server at the address with HTTP transport and set the header with the token.
  • If the client cannot pass a header, use the address with the token inside: …/mcp/<token>. Keep such an address like a password — it opens access to the project.

After connecting, ask the assistant: “What flows does my bot have?” — if it lists them, everything works.

Example requests

  • “How many new leads did the bot bring this week and from which channels?”
  • “Show conversations where the customer is waiting for an operator and briefly summarise each”.
  • “Find the customer who asked about delivery to Kazan and show the conversation”.
  • “Compare the number of conversations for the last 7 and 30 days”.
  • “Reply in conversation 1520: the order has been passed to delivery, we will send the tracking number in the evening” — the assistant will call send_reply.
  • “Close the spam conversations from the list”.

The assistant works with the same data as the dashboard — it just saves clicks and can summarise and compare right away.

Replies to customers: how it works

The send_reply tool behaves exactly like an operator reply from Conversations: the message goes to the customer in the messenger they wrote from, is added to the chat history, and the conversation moves to an operator — the bot goes silent in it.

The tool description tells the assistant to show you the text and get your consent before sending. Good practice is still to ask for a draft first and only then confirm sending.

Rights and security

The CHBT MCP server uses the same rules as the public API: the project token, a request limit and project isolation — the assistant sees only your bot.

  • Every token gives the full set of tools. A token from API connections allows both reading data and replying to customers. So issue a separate token for each person and each assistant, and revoke an employee’s token when they leave.
  • Replies only with consent. The send_reply tool description asks the AI to show you the text and get your consent before sending. Do not turn off action confirmation in your assistant.
  • Revoke in a second. In API tokens press Withdraw — the assistant loses access immediately.
  • A header is safer than an address with a token. An address with a token can end up in proxy and browser logs; use it only if the client does not support headers.
  • Personal data. The assistant receives customers’ conversations and contacts. Connect only an AI service you trust with this data, and keep the requirements of Federal Law 152 in mind.

MCP, API or webhooks — which to choose

  • MCP — when a person asks the questions through an AI assistant: the owner, a marketer, a shift lead.
  • Public API — when a developer writes the integration: your own CRM, helpdesk, reports.
  • Webhooks — when an external system should learn about events by itself: a new lead, a new conversation. For amoCRM and GetCourse CHBT has ready-made integrations — leads go there with no code.

All three use the same tokens from API connections. Step-by-step setup is in the help article “MCP: the CHBT bot in Claude, ChatGPT and Cursor”, an overview is on the integrations page and in the article on incoming webhooks.

Common mistakes

  • One token for everyone. When an employee leaves, you have to revoke access for everyone. Issue a token per person.
  • A token for an assistant you do not trust. Any token allows replying to customers — give it only to those you trust with your conversations.
  • An address with a token in a group chat. It is a password to the project. If it leaked, revoke the token and issue a new one.
  • Replying to a customer without checking. Ask for a draft and confirm sending.

Use case: a morning review in five minutes

What an ordinary owner’s morning with MCP can look like:

  1. “How many leads came in yesterday and from which channels?” — the assistant calls get_stats and list_leads and answers with a list of phone numbers.
  2. “Which conversations are waiting for an operator?” — list_conversations with the “waiting for an operator” status. The assistant briefly summarises each.
  3. “In the conversation with Marina, reply that the master is free at 15:00, show a draft” — the assistant prepares the text, you confirm, it calls send_reply.
  4. “Close conversations where the customer said thanks and left” — close_conversation for the list you checked.

All this without logging into the dashboard, from the window where you already work with AI.

Who on the team will find MCP useful

  • The owner — quick numbers: leads, conversations, channels, period comparisons.
  • A marketer — where leads come from and which questions are asked most often, without exports.
  • A shift lead — the list of conversations waiting for an operator and a short summary of each before assigning.
  • A developer — checking that an integration works and viewing raw conversation data in Cursor or Claude Code.

Everyone gets their own token: this shows who uses what, and anyone’s access can be revoked separately.

What to do if the assistant does not see the tools

  1. Check the address: it must end with /mcp and point to your dashboard.
  2. Check the header: Authorization: Bearer and the token without extra spaces.
  3. Make sure the token has not been revoked — API tokens show its state and when it was last used.
  4. If the token was lost or revoked, issue a new one and update it in the assistant settings.

Integrations CHBTIncoming webhook in a chatbotHow to set up an AI agentHelp: connecting MCP

More on the blog