// insights
How to connect your business data to ChatGPT and Claude with MCP
A growing share of your customers now ask an AI assistant before they ask a search engine — and those assistants can't see your product catalog, your inventory, or your order data unless you connect it. MCP is how you do that: one server that lets ChatGPT and Claude read your real data and, when you allow it, take action on a customer's behalf.
The problem: your business doesn't exist to AI yet
Ask ChatGPT or Claude a specific question about your products, your pricing, or your availability, and it either says nothing useful or guesses from whatever happens to be indexed online. Your real data — catalog, inventory, order status, account history — isn't something either assistant can see. A competitor who gets connected first becomes the one AI recommends, cites, and quotes.
What MCP actually is
The Model Context Protocol (MCP) is an open standard for connecting AI assistants to real data and tools through a permissioned interface. Instead of building a separate chatbot that only lives on your own website, you build one MCP server — exposing exactly the product specs, inventory, or account actions you choose to share — and any AI assistant that speaks MCP can use it.
One server, plugged into both ChatGPT and Claude
This is the part that makes MCP worth building on: it isn't a custom chatbot locked to your own site. It's a standard both major AI assistants already speak.
Claude
Anthropic's Claude — on claude.ai, Claude for Work, and Claude Desktop — connects to remote MCP servers through its Integrations feature. Once your server is live, anyone researching in Claude, or a procurement team using Claude for Work, can pull accurate answers straight from it.
ChatGPT
OpenAI has adopted the same open MCP standard, letting people and organizations connect approved MCP servers as Connectors inside ChatGPT. The server you built once now reaches ChatGPT's audience too — no second integration.
Build it once, and it becomes a live presence inside both — the two AI assistants increasingly sitting between your customers and Google.
It doesn't just answer — it can take action
This is where MCP goes further than a search-and-answer chatbot. Each capability you expose is a specific, named tool — some let the assistant read data, others let it actually do something on a customer's behalf. You decide which of your systems get a read tool, a write tool, or none at all.
Read
- Check current order or shipment status
- Look up product specs, availability, or certifications
- Pull current inventory or pricing
Act
- Create a quote or order request
- Update a record in your CRM
- Schedule a callback or site visit
- Open a support ticket
Every action tool is scoped to exactly what it's allowed to change, requires explicit intent before it runs, and is logged — the same discipline we'd build into any production system, applied to what an AI is allowed to do on your behalf.
What about the security of your data?
Giving an AI assistant access to real business systems raises an obvious question: what stops it from seeing — or doing — something it shouldn't? MCP is built around that exact concern, and it's the part of this we take most seriously.
The model never touches your systems directly
No AI assistant ever gets your database credentials or API keys. It can only call the specific tools your MCP server exposes — each one narrow, named, and fixed in what it's allowed to do.
Every connection is authenticated
Your server only responds to approved assistants and permissioned users — not anyone who happens to find its address. Customers, staff, and partners can be scoped to different access levels.
Least privilege by default
Customer-facing tools return customer-facing data. Margins, internal notes, and other customers' records are never exposed in the first place — not filtered out after the fact.
A full audit trail
Every read and every action is logged — what was accessed, what was changed, and in which conversation — so you can review, and reverse, anything an assistant did on your behalf.
In practice, your data is more controlled with MCP than it is today — instead of scattered API keys and one-off integrations of varying quality, everything an AI can read or change flows through one server you built, permissioned, and can audit.
What this unlocks
Show up where the research already happens
Buyers increasingly ask ChatGPT or Claude before they Google. If you're not connected, you're not part of that conversation — and a competitor who got there first is.
Be the accurate answer, not a guess
Without MCP, an AI assistant asked about you either says nothing or guesses from outdated, third-party info. With it, it answers from your real, current data.
Cut repetitive support and sales questions
The same spec, status, or pricing questions your team answers by phone or email can be answered — and acted on — directly inside the conversation, through tools you control.
Stay open 24/7 without adding headcount
An MCP server doesn't take weekends off. It answers the question that comes in at 2am the same way it answers the one at 2pm.
Case study
What this looks like in practice
We used AWA Roofing — a climate-resilient steel roofing manufacturer in Ghana, and one of our own clients — as a concrete example while scoping this. Here's what an MCP server would put in front of ChatGPT and Claude users on their behalf:
- 01
A contractor needs a spec answered — and a quote started
They ask Claude which steel roofing profile handles heat gain and heavy rain best. Claude calls AWA Roofing's MCP server, pulls the real Cool Roof Technology and Snaplock specs, then — with the contractor's go-ahead — calls a request_quote tool right there in the conversation. No form, no separate email.
- 02
A customer abroad is comparing suppliers
From overseas, they ask ChatGPT to compare steel roofing manufacturers in Ghana. Connected to AWA Roofing's MCP server, it answers with current, verified specs and certifications instead of outdated third-party listings.
- 03
A procurement team is shortlisting certified suppliers
Their company uses Claude for Work internally. Asked to shortlist certified IBR suppliers, it surfaces AWA Roofing because the company is a connected, structured source — not an invisible one.
Is this worth building for you?
MCP is usually worth scoping if two or more of these are true:
- Customers or prospects research you — or your competitors — before they buy
- Your team fields the same spec, status, or pricing questions on repeat
- Your product, inventory, or account data lives in a system an AI can't currently see
- You've already got an AI agent or chatbot project somewhere on the roadmap
How we build MCP servers
MCP development is one of our four core services, alongside website, mobile, and AI agent development. We scope exactly what the server can read and what it can act on — pricing and inventory data never leave your systems ungoverned — then ship it as production infrastructure, not a prototype. The same weekly-release process we use for every engagement applies here: working software every week, observability from day one.
Ready to be findable by AI?
Tell us what you sell and where your customers already ask questions. We'll scope what an MCP server for your business would look like — free, no commitment.
Start a project