Claude and ChatGPT connectors
MCP integration: are you ready for ChatGPT and Claude?
Your users will ask an AI assistant about your product either way. The only choice is whether the answer comes from you.
Short answer
An MCP integration lets ChatGPT and Claude answer your users' questions from your product, under each user's own login and your rules. A product is ready when it has a working API, a sign-in that knows each user and customer, permissions checked on every record, a staging system with two test users, and a clear list of fields an assistant may show.
On this page
- 01Why product companies are asking about this now
- 02Five readiness checks before an MCP integration
- 03Pick your first three tools from real questions
- 04What a read-only connector looks like, layer by layer
- 05How to test it before anyone uses it
- 06What we see in production
- 07Where to go next
- Frequently asked questions
01Why product companies are asking about this now
The people who use your product already work in ChatGPT and Claude for part of the day. They paste exports into a chat, ask about their accounts and copy answers back. When that happens, the assistant answers from whatever the user pasted, with no view of your permissions and no way to check what is current.
MCP, the Model Context Protocol, is the open standard AI assistants use to reach other software. ChatGPT and Claude both connect to outside products through it, so one well-built MCP server can serve both. That changes the question for a product company. It is no longer "should we build an AI assistant?" It is "when our users ask an assistant about their work, does the answer come from our product, under our rules?"
For most established products the honest answer today is no, and the first gap is rarely the protocol. It is everything behind it.
02Five readiness checks before an MCP integration
We run the same five checks before any connector work starts. Each one is a yes or no, and a single no is a reason to fix the product first.
| Check | What ready looks like | Why it matters for the connector |
|---|---|---|
| A working API for the tasks users ask about | Documented, and used by your own screens | The connector calls your API. If your own screens do not use it, it will drift from what users see. |
| A sign-in that knows each user and each customer | The customer comes from the login, never from the chat | An assistant can be asked anything. The customer the answer is for must come from a verified token. |
| Permissions checked on every record | In the backend, not only hidden in the screens | A field hidden in the interface but returned by the API will reach the chat. |
| A staging system with two or more test users | Separate test users, ideally in separate customers | You need to prove that one user cannot see another user's data before anyone real tries. |
| Data you are happy for an assistant to show | A field list, picked one field at a time | Whatever a tool returns lands in the user's chat, and from there it can be copied anywhere. |
The third check catches most teams. Products that grew over years often enforce some rules only in the screens: a column that is never rendered, a button that is hidden for one role. A connector does not use your screens. It reads what the API returns, so the API has to apply the rules itself. We go deeper on this in tenant isolation.
03Pick your first three tools from real questions
Once the five checks pass, write down the three questions your users ask again and again. Look at support tickets, sales calls and the reports people export most. Those three questions become your first three tools.
Each tool gets a one-page contract before any code is written. Here is an illustrative one for a fictional invoicing product:
| Field | Contract |
|---|---|
| Name | find_overdue_invoices |
| Does | Lists invoices more than N days overdue, for the companies this user can see |
| Asks for | days_over, from 1 to 365. Nothing else. |
| Calls | One known API request, made as the signed-in user |
| Returns | Customer, invoice number, days overdue. No contacts, no notes. |
| Limit | 25 rows at most, oldest first |
| Who may use it | Finance and admin roles. Every other role is denied. |
A contract like this forces the decisions that matter: which fields leave the product, how many rows, and which roles can ask at all. It is also the document a security reviewer will want to see later.
Start read-only. A tool that changes data waits until the read-only ones have earned trust, and then asks the user to confirm every change.
04What a read-only connector looks like, layer by layer
The connector is the only new part. Your login, your API and your data keep working exactly as they do today.
- 01
The assistant.
ChatGPT or Claude. The user asks a question, and the assistant picks a tool.
- 02
The sign-in.
Your existing login, through the authorization flow the MCP standard sets out. The connector accepts only tokens issued for it.
- 03
The connector.
A few tools, running in your cloud. It checks who is asking, which tools that person may call, and logs every call.
- 04
Your API.
Same rules, same permissions. It checks the user's rights again on every call, as it would for your own screens.
Some things never go through a connector, at least at the start: delete and bulk export endpoints, admin endpoints, other customers' records, and any field you chose to leave out. Leaving them out of the tool list is stronger than telling the assistant not to use them, because a tool that does not exist cannot be talked into anything.
05How to test it before anyone uses it
A connector is a new door into your data, so it gets tested like one. In staging, with at least two test users in two different customers, we check that:
- One user asks for another user's record by name, by ID and by search, and is denied each time.
- A user without the right role cannot call a tool at all.
- A tool returns only the fields in its contract, even when the record holds more.
- Row limits hold when the user asks for "all of them".
- An expired or revoked login stops the connector at once.
- Every call appears in the log with the user, the customer and the tool.
We hold every connector to six access tests like these in staging before anyone uses it, as we describe in our guide to putting your product inside ChatGPT safely. The same tests then run on every release, because a new API field or a new role can open a gap that was closed last month.
06What we see in production
We run our own operations this way. Our operations system is live in ChatGPT and Claude and in daily use. Each person signs in with their own account, keeps their own role, and can make only the updates that role allows.
For CleanHaus, we built an agent that works through CleanHaus's own APIs, with a tool list per role and each user's own login. The agent has no permissions of its own. That pattern, the assistant acting as the signed-in user through the APIs that already exist, is what keeps an MCP integration small. You are not building a second product. You are opening a narrow, logged door into the one you have.
The most common blocker we find is the third readiness check. A team wants to connect, the API returns the full record, and the screens were doing the filtering. Fixing that is worth doing whether or not a connector ever ships, because the same gap is open to any other integration.
07Where to go next
If the five checks pass, the build itself is the smaller part. We cover the server in how to build an MCP server for your product, and the differences between assistants in our guides to the ChatGPT connector and the Claude connector.
To run the five checks on your own product, use our free field guide, Will your product work inside ChatGPT and Claude?. If connecting comes first, our MCP server development service, Tenon, starts with up to three read-only tools your users need most, each user signed in with their own account, tested in your staging before anyone uses it.
Frequently asked questions
What is an MCP integration?
It is a small server, built on the Model Context Protocol, that lets AI assistants such as ChatGPT and Claude call your product's tools on behalf of a signed-in user. The assistant picks a tool, the server checks who is asking and what they may do, and your API returns only what that user is allowed to see.
Is our product ready for ChatGPT and Claude?
It is ready when five things are true: a working API for the tasks users ask about, a sign-in that knows each user and each customer, permissions checked on every record in the backend, a staging system with two or more test users, and a list of the fields an assistant may show. If any one of them is missing, fix that first.
Can one MCP server work in both ChatGPT and Claude?
Yes. Both assistants connect to outside products through MCP, so one well-built server can serve both. How each assistant is turned on for a company, and which admin controls the customer gets, differs between the two, so test the connector in each one before your users do.
Should the first tools be able to change data?
No. Start with read-only tools for the three questions users ask most. Add a tool that changes data only after the read-only ones have earned trust, and have it ask the user to confirm every change before it runs.
How Datics helps
Datics specializes in Brownfield AI: production AI inside software people already rely on. Tenon Starter puts your product inside ChatGPT and Claude with up to three read-only tools, signed in with each user's own account, for $2,500.
More on Claude and ChatGPT connectors
Claude and ChatGPT connectors
How do you build a Claude connector for your product?
A Claude connector lets your users ask Claude, or ChatGPT, about their data and get answers from your product, under their own login and your rules.
Claude and ChatGPT connectors
ChatGPT connector: put your product inside ChatGPT safely
A ChatGPT connector is an MCP server that lets users reach your product from ChatGPT under their own login. How it works, what to secure and when to wait.
Claude and ChatGPT connectors
Build an MCP server for your product: a production guide
Build an MCP server as a thin remote service over your API: a few tools from real questions, Streamable HTTP, OAuth per user, limits, logs and real tests.
Your customers already expect AI from your product.
Let us make it real inside the product you run today.
