ChatGPT MCP for Your AI Agent, Both Ways

Search for ChatGPT MCP and you get setup guides for developers. This one is for the person who runs the business website and its AI agent. MCP is useful to you in two directions: you can run your agent from the AI app you already work in, and your agent can work in the booking, order or payment system you already pay for.

What does MCP actually mean?

MCP (Model Context Protocol) is an open standard that lets AI apps such as ChatGPT and Claude connect to outside software: they read its data and use its actions as tools. The project's own site, modelcontextprotocol.io, compares it to a USB-C port for AI applications (read November 4, 2026). Anthropic open-sourced it on November 25, 2024; OpenAI, Microsoft and Cursor build it into their tools.

Two roles make it work. An MCP server sits in front of a piece of software and lists what an AI may do there, such as "find open slots" or "list leads". An MCP client is the AI app that connects, reads that list and calls the tools when a conversation needs them. One server serves every client, so a business system with an MCP server works with ChatGPT, Claude and the next AI app alike.

ChatGPT MCP for your agent works in two directions

For a business with a website AI agent, MCP runs both ways. In the first, ChatGPT or Claude is the client and Ultimo Bots is the server: you manage your agent from the chat. In the second, your agent is the client and a system you already use is the server: the agent looks things up and books there while it talks to a visitor.

YouYou askUltimo Bots
MCP serverYour agentA visitor asksMakeplans
MCP serverBookings

Ultimo Bots is the AI agent platform where you describe in plain words what your agent should do, and Mo, the builder in the AI Studio, builds it: skills on public APIs, on remote MCP servers and on ready-made app actions, so the agent does real work in your own systems. It also runs its own MCP server, which is where direction one starts.

Direction one: connect ChatGPT or Claude to your agent

The Ultimo Bots MCP server lives at https://mcp.ultimo-bots.com/mcp, and you add it to your AI app like any custom MCP server. You sign in with your Ultimo Bots account, approve what the app may touch, and from then on you ask in plain language: "What did visitors ask this week?" or "Change the welcome message."

Both apps use the same URL and an OAuth sign-in on your Ultimo Bots consent page. In ChatGPT the path follows OpenAI's guide Connect and test your plugin: enter the URL as a public endpoint, set the authentication to OAuth (the Ultimo Bots server uses it), confirm the warning, and select Create as a plugin. In Claude it follows Anthropic's help article on custom connectors. Both guides were read November 4, 2026. Codex and Claude Code connect the same way; the portal's MCP connections page shows the server URL and a ready prompt for both.

The consent page decides what the AI app may do

After the sign-in, Ultimo Bots shows "Connect an AI assistant". You choose one workspace, then Only selected agents or All agents in this workspace, including future agents, then the permissions, from "Read conversations" and "Read leads" to "Edit widget appearance" and "Edit business context". Write permissions need their read permission, and your own role still applies, so the AI app can never do more than you could in the portal. Click Allow connection and you are back in the chat.

The connection stays active while you use it and ends after 30 days without use, after a year at the latest. On the MCP connections page in the portal you see every connected app and its agents, and Revoke access cuts it off at once.

What you can ask for once it is connected

• Conversations: list and read what visitors asked, filtered by date or channel.
• Leads: see new requests, mark them read, fill in fields.
• Insights: the dashboard figures for a date range you pick, plus your Success Score.
• Widget: welcome messages, starter questions, colors, header text, position.
• Business context: read it and replace it, up to 1,000 characters.
• Live chat and notifications: live chat settings, automatic replies on or off, report and alert settings.

Take Hemline Alterations, a fictional tailor and alterations shop. On a Monday morning Rosa, the owner, has Claude open for her invoices anyway, with Ultimo Bots connected for her one agent. She asks about last week's chats and any new leads. Claude reads the chats and the two new leads and finds five questions about wedding dresses that her business context does not cover.

Managing your agent this way uses no AI credits: ChatGPT or Claude does the thinking on your plan with them. AI credits pay for your agent's answers to visitors and for Mo. Skills and behaviors take shape in the AI Studio, where Mo builds and tests them with you.

Can I use MCP with OpenAI?

Yes. ChatGPT connects to remote MCP servers, and OpenAI packages them as plugins. A plugin bundles an MCP server connection with its sign-in, sometimes with skills; on ChatGPT's Plugins page you can install published ones or add a custom MCP server by URL, which is what you do for Ultimo Bots. OpenAI's Codex, in the terminal and in the ChatGPT desktop app, reads MCP servers from its own settings.

Which ChatGPT plans may add custom MCP servers is OpenAI's call, so check OpenAI's help center before you start. In a Business or Enterprise workspace, the workspace's policies decide whether members may add one. Once the plugin is added, the Ultimo Bots tools appear in ChatGPT like any other plugin's.

Direction two: your agent books in the system you already use

A growing list of business tools publish official MCP servers, Stripe and Makeplans among them, and your Ultimo Bots agent can use one as a skill. Mo connects to any remote MCP server on a public HTTPS address that takes an API key or bearer token, reads the list of tools it offers, and builds a skill around the ones your job needs.

Back at Hemline, fittings are booked in Makeplans, a booking system for appointments. Its MCP guide (updated September 30, 2026, read November 4, 2026) describes a server at https://<account>.makeplans.com/api/mcp over stateless Streamable HTTP, with the account's API key sent as a bearer token once the API is switched on. It offers eight tools, among them list_services, find_available_slots and create_booking. Rosa opens the AI Studio and types, or dictates, her message to Mo:

1. 1

   Mo reads the vendor's docs

   It looks up the official MCP server first and learns which tools exist and what each one needs.

2. 2

   You paste the API key

   Mo asks for the Makeplans key in a secure field. It is stored encrypted and never sent to the AI.

3. 3

   Mo connects and builds

   It connects to the server, picks the tools for this job and builds the skill. Booking asks the visitor to confirm the summarized details first.

4. 4

   Mo tests and reports

   It tries each tool, the booking as a dry run, then runs a short test chat that calls Makeplans for real, and tells you what the agent can now do.

On a Wednesday afternoon a visitor opens the chat on Hemline's website; a behavior makes the agent introduce itself as Hemline's AI assistant. The visitor writes: "Can I get a fitting for a bridesmaid dress this week?" The agent replies: "I'm Hemline's AI assistant. Free fitting slots this week: Thursday at 11:00 am, Thursday at 4:30 pm and Saturday at 10:00 am. Which one suits you?"

Priya picks Thursday at 4:30 and gives her name, email and phone. The agent replies: "Bridesmaid dress fitting on Thursday at 4:30 pm for Priya Shah, priya@example.com, 555‑0142. Shall I book it?" After her yes it books the slot in Makeplans and confirms the time.

The slots are invented here; in a real chat they come from Makeplans at that moment. Two practical notes for your own setup. Test chats call Makeplans for real, Mo's short test chat and your own in the Test tab alike, so a test booking lands in your calendar: cancel it in Makeplans afterwards. And from Boost up, with Voice input switched on in the Design tab, visitors can also speak their question into the widget; the agent answers in writing.

The same pattern fits every system whose official remote MCP server takes an API key: a shop, a ticketing tool, an events calendar. Where a system has no MCP server, Mo builds the skill on its public API instead, and for webhooks there are Zapier, Make and n8n. The building blocks are on the AI Studio page and the ready-made ones under app actions.

Each MCP skill counts as one of your plan's skills, and the agent's answers and Mo's work use AI credits from one balance per agent: Smart $29 a month ($19 billed yearly) with 500 AI credits and 3 skills, Boost $59 ($39) with 2,000 and 10 skills, Ultimo $249 ($199) with 10,000 and 30 skills. A quick question to Mo uses a few credits, building a skill more (pricing).

Which AI supports MCP?

Major AI assistants and coding tools support MCP as clients: ChatGPT and Codex from OpenAI, Claude and Claude Code from Anthropic, Cursor and Visual Studio Code. Each adds remote servers by URL; the menus differ, the protocol does not. That is why one Ultimo Bots MCP server works in all of them, and why your agent can use the same Makeplans server that Rosa could also connect to Claude for herself.

Your Ultimo Bots agent is an MCP client too, with one difference: it connects only to the servers Mo set up for its skills, with the keys you stored, and calls only the tools those skills include. A visitor's chat never adds a server on its own.

Is MCP just tool calling?

No. Tool calling is the AI model's ability to ask for an action; MCP is the shared plug that tells any AI app which actions a system offers and how to call them. Before MCP, every AI app needed its own integration with every system. With it, the system describes its tools once, and ChatGPT, Claude or your agent read the same description.

For a business that means less lock-in. The connection to your booking system belongs to the system, not to one AI vendor, and moving from one AI app to another does not mean rebuilding it. MCP also covers more than tools, such as documents a server shares and ready prompts, but tools are what both directions in this guide use.

Test it with one read-only request

The safest first test of ChatGPT MCP is a question that changes nothing. Add https://mcp.ultimo-bots.com/mcp in ChatGPT or Claude, leave only read permissions ticked on the consent page (workspace, agents, conversations, leads, reports and insights), and ask: "What did visitors ask my agent this week?" The answer shows you, in your own data, what your visitors want and what the agent did not know.

Then pick the system your team opens most often to answer customers. If its vendor lists an official MCP server, send Mo that page and one sentence about the job; if not, send the API docs. Check new leads in leads and analytics and switch yourself online in live chat for the questions that still need a person.

Questions and answers