ChatGPT and Claude are already useful to estate agents without any setup: a pasted enquiry becomes a clear reply, a messy set of notes becomes a viewing summary. The bigger step is connecting them to your CRM, so the assistant answers from your own pipeline instead of from whatever you remember to paste in. That connection runs on MCP, and this guide explains what it is, what to ask, what never to paste, and where the limits are. It is part of our wider guide to AI for estate agents.
What ChatGPT and Claude do well for estate agents
General chatbots are strongest at language work on text you give them. In an estate agency that means:
- Drafting. Replies to enquiries, viewing follow-ups, vendor updates, chasing emails.
- Summarising. Turning a long email thread or a page of call notes into five lines and a next step.
- Rewriting. Making a listing description shorter, warmer, or fit for a different portal.
- Organising. Sorting a list of tasks by urgency, or preparing questions before a valuation.
They are weaker at anything that needs facts they cannot see: your live pipeline, the price you agreed last Tuesday, which buyer asked for a garden. Without a connection, you are the courier between your data and the chat window. That is slow, and it is how personal data ends up where it should not.
What MCP is and why it matters here
According to the official MCP site, MCP (Model Context Protocol) is an open-source standard for connecting AI applications to external systems, so that assistants such as Claude or ChatGPT can reach data sources, tools and workflows. The site compares it to a USB-C port for AI applications: one standard way to plug things in.
For an estate agent, the practical meaning is simple. A CRM publishes an MCP server. You add that server to Claude or ChatGPT as a connector. From then on you can ask the assistant a question about your leads and it fetches the answer from the CRM, instead of guessing.
What the plans require, as of 20 October 2026
Plan rules change often, so check the vendors' own pages before you commit. This is what they say today.
- Claude. Anthropic's help article on custom connectors using remote MCP lists the Free, Pro, Max, Team and Enterprise plans. Free accounts are limited to one custom connector. On Team plans only Owners and Primary Owners can add them.
- ChatGPT. Which ChatGPT plans can add a custom MCP connector, and whether it can change data or only read it, is set by OpenAI and is still in beta, so it changes. Read OpenAI's help article Developer mode and MCP apps in ChatGPT and Formule's connect ChatGPT guide for the current rules. On workspace plans an admin or owner controls access.
If you are on a team plan, the person who can add the connector is often the owner or an admin, not you. Ask them early.
Other CRMs offer AI connectors too
Formule is not the only CRM with an MCP connector. onOffice, the German CRM, has an AI Connector. Its page says it links onOffice enterprise with MCP-compatible AI tools, covers properties, contacts, appointments, activities and tasks, and is included in the onOffice enterprise all-in and expert plans. It names ChatGPT, Claude, Mistral and Langdock, and says it works strictly within the rights of the logged-in user. The same page tells users that AI can make mistakes and that results should be verified, which is advice worth following with any connector.
Whichever CRM you use, ask the vendor three questions: whose permissions does the connection run under, which actions can the assistant take that change data, and where does the data go when the assistant reads it.
Nine prompts an agent can use
The first five assume a connected CRM. The last four work in any chat if you paste in text with the personal details removed. Replace the square brackets with your own details.
Pipeline questions
- Uncalled leads. "Which leads came in this week that nobody has called? List them newest first with the property each one asked about."
- Stale leads. "Show me open leads with no note in the last 10 days and suggest one next step for each."
- Matching. "Which of my leads fit the new [two-bedroom flat] in [neighbourhood] at [price]? Give a one-line reason for each."
- Month on month. "How is my pipeline doing compared with last month? Say what changed, in plain words."
- Before a call. "Give me everything on [lead name]: what they asked for, what I promised, and what is still open."
Drafts and summaries
- Follow-up after a viewing. "Write a short, friendly follow-up to a buyer who viewed [a three-bedroom house] yesterday and liked the kitchen but worried about the road noise. Offer a second viewing. Under 80 words."
- Reply to a portal enquiry. "Here is an enquiry with the name removed: [text]. Draft a reply that answers the question, asks one qualifying question and suggests two viewing slots."
- Listing summary. "Turn these facts into a 120 word listing description and a three line summary for a message: [bedrooms, size, features, condition]. Use only the facts I gave you."
- Vendor update. "Write a weekly update for a seller: [number] enquiries, [number] viewings, feedback was [summary]. Keep it honest and suggest one change to try."
The pattern behind all of them: give the assistant the facts, say who the message is for, set a length, and tell it to use only what you supplied.
What never to paste into a chatbot
A consumer chat window is not a place to put client files. Do not paste:
- Names, phone numbers, email addresses or home addresses of clients, unless you have a lawful basis to share them with that provider and have checked its terms.
- Identity documents, proof of funds, bank details or anything from anti-money-laundering checks.
- Financial or health details a buyer or tenant told you in confidence.
- Another agency's confidential material.
The safer habit is to remove names and contact details before you paste, or to use a connection that runs under your own permissions so the data is not copied around by hand. Data protection rules apply to both routes, and the point is covered in our guide to GDPR and the EU AI Act for estate agents. If your agency has a policy on AI tools, follow it.
Limits you have to plan for
- Hallucinations. Assistants produce fluent text even when they are wrong. A listing description may invent a feature, a summary may drop a condition. Tell the assistant to use only the facts you gave, then check the result against the source.
- Check before sending. Read every message before it goes to a client. Names, dates, prices and promises are yours, not the model's.
- Stale or partial data. A connector shows what the CRM holds. If your notes are thin, the answers will be thin.
- Actions that change data. Be careful with any connector that can add notes, change statuses or send messages. Confirm each change before it runs.
- Plan and admin limits. Features depend on your plan and on what an admin allows, as the plan section above shows.
How Formule connects to Claude and ChatGPT
Formule is an estate agent CRM with agentic AI: Really, its AI agent, does the work a CRM usually only records. Separately from Really, Formule runs an MCP server so you can use Claude or ChatGPT on your Formule data. This is what the Formule MCP documentation says.
- It runs as you. Every tool call uses your own login and the same row-level security as the app. The docs state that whatever you can see in the app is what the assistant can see, no more. What you can see depends on your role and your agency's access settings.
- Address and sign-in. The server is at mcp.formule.io. Sign-in uses OAuth, and you approve the connection on app.formule.io as yourself.
- Claude. In Claude.ai or Claude Desktop you open Settings, then Connectors, add a custom connector with the Formule address and approve. A test question such as "Who am I in Formule?" should return your name, role and agency. The docs say Claude Code is not supported yet.
- ChatGPT. ChatGPT needs the address ending in /mcp and works on the web, not in the phone or desktop apps. Which plans get full access, including changes such as adding a note (which OpenAI labels beta), and which are read-only depends on OpenAI, so follow the plan notes in Formule's ChatGPT guide. Before ChatGPT changes anything in Formule it asks you to confirm and shows what it will send.
- What the tools do. According to the docs there are 30 tools. They read leads, listings, property matches and your pipeline stats, and they can add notes, change statuses and schedule follow-ups. Read-only tools are marked as such, so a client can tell which calls are safe. A tool that sends a WhatsApp message is listed too, and the docs say it is only to be used after you approve the exact text.
Formule's docs list prompts such as "What leads came in this week that nobody has called?" and "Which of my leads match the new two-bed in the centre?", which are the first and third prompts above. If you would rather stay inside Formule, Really performs similar tasks in the app, on WhatsApp and by voice, in English, German, French, Spanish, Italian or Polish, and asks before making a change. For follow-up with WhatsApp specifically, see our guide to AI lead follow-up on WhatsApp. To turn a property link into a listing draft, see AI property listing from a link.
A sensible way to start this week
- Pick one assistant and one task, for example post-viewing follow-ups.
- Write a standing instruction with your tone, your language and your sign-off, and reuse it.
- Run the prompts above on text with the personal details removed, and note what you had to correct.
- Connect your CRM through MCP only once you know who is allowed to add it on your plan, and test with a question about yourself before a question about your leads.
- Keep the rule that nothing goes to a client until you have read it.
You can see how Formule works for UK agents on the Formule UK page, or start free and let Really set up your account in one chat.