Query your company data from ChatGPT or Claude

MCP, explained in two minutes

The official Model Context Protocol documentation defines it as an open-source standard for connecting AI applications to external systems, and compares it to a USB-C port: a single socket, whatever device is plugged in.

Published by Anthropic in November 2024, MCP is now supported by many assistants, including Claude and ChatGPT. In practice, your software exposes a list of “tools”: standard questions such as “unpaid invoices” or “today’s jobs”. The assistant chooses the tool that matches your question, your software sends back the answer, and the assistant formats it.

01
You ask the question
In Claude, ChatGPT or your app’s chat, in plain language.
02
The assistant chooses the tool
It calls the matching connector tool, for example “pending quotes”.
03
Your software checks and answers
It checks that the user is authorised, sends back the requested data read-only and logs the access.
04
The assistant writes the answer
It presents the result, summarises it or compares it. You can ask follow-up questions.

Sources: modelcontextprotocol.io and Anthropic’s announcement, accessed 4 October 2026.

What you can ask your software

Questions your teams ask a colleague, a spreadsheet or several software packages at once today.

Management

“Where does site X stand?”
“How has revenue changed since the start of the year?”
“Which alerts do I need to handle today?”

Sales

“Which quotes have been waiting for an answer for more than 15 days?”
“What is the quote conversion rate this quarter?”
“What has happened at this address over the past two years?”

Scheduling and field

“What is planned on my site tomorrow?”
“Which jobs are not assigned to anyone yet?”
“Which job reports have not been done?”

Accounting and cash

“Which customers have unpaid invoices?”
“Which payments did we receive this week?”
“What is our cash position?”

Cross-cutting questions are the most useful: “Which customers have both a pending quote and an unpaid invoice?”. They assume that all this data lives in the same place — which is where the connector of your own software makes the difference.

How C-Esium delivers the connector, from V1

The MCP connector is part of every internal software built by C-Esium. Its tools follow your data model — customers, quotes, jobs, invoices — and the questions your teams actually ask.

Read-only

The connector reads, it never writes: nothing can be created, changed or deleted from the assistant.

Authorised users

You decide who can query the software from an AI assistant. Other users have no access.

Logged

Every access is logged, so you can review how the connector is used afterwards.

Claude, ChatGPT or the app

The same connector serves the market assistants and the copilot built into your software. Like the rest of the project code, it is assigned to you at the end of the project.

The MCP connector of an off-the-shelf tool, or the one of your own software

Vendors now publish their own MCP connector: HubSpot for its CRM, for example, listed in Claude’s connector directory. If most of your business lives in one such tool, plug in its connector: it is simple and useful. The difference appears when your business is spread across several tools, or when your processes fit into none of them.

Source: Claude connector directory (HubSpot), accessed 4 October 2026.
Criterion	Connector of an off-the-shelf tool	Connector of your internal software
Data covered	The vendor’s scope: its CRM, its bank, its tool.	All the data in your software: requests, quotes, jobs, scheduling, purchasing, invoicing, cash.
Data model	The vendor’s, the same for all its clients.	Yours: your business objects, your statuses, your rules.
Cross-cutting questions	Possible by combining several connectors, if the information matches from one tool to the next.	Native: a customer, their quotes, jobs and invoices are in the same database.
Tools available	Defined by the vendor, according to its roadmap.	Defined with you, from your teams’ questions, and changeable.
Read and write	Depending on the vendor and the administrator’s settings.	Read-only, logged, authorised users.
Ownership	The connector belongs to the vendor.	The code is assigned to you with the rest of the software.

What leaves your software, and what does not

Answers pass through the vendor of the assistant you use. Knowing exactly what leaves your software helps you choose the right plan and frame the contract.

Which AI plan should you choose?

C-Esium recommends a business plan: it puts use under a company contract and, according to the vendors, does not use your data to train the models by default. The framework chosen is set out in your contract with C-Esium.

Sources: OpenAI (learn.chatgpt.com, developers.openai.com) and Anthropic (claude.com, privacy.claude.com, support.claude.com), accessed 4 October 2026. Plans change: check the terms when you subscribe.
Plan	For whom	Model training	MCP connectors
ChatGPT Business	SME teams	No training on information obtained through connectors, according to OpenAI.	The administrator chooses which roles can use or install a connector, and can limit compatible connections to read actions.
ChatGPT Enterprise	Mid-sized companies, advanced control needs	No training on company data by default, according to OpenAI.	Same controls: the administrator enables MCP connections, assigns them by role and can limit them to read actions.
Claude Team	SME teams	No training on your content by default, according to Anthropic.	Added by the organisation’s owners, then connected individually by each member.
Claude Enterprise	Mid-sized companies, advanced control needs	No training on your content by default, according to Anthropic.	Added by the owners or an authorised role, then connected individually by each member.
OpenAI or Anthropic API	The chat built into your app	No training by default, according to both vendors.	The connector is called by your own application.
Individual accounts (free or paid)	To avoid for company data	Consumer terms, outside a company contract.	No central management by the company.

Data protection: three points to check

Querying your data from an AI assistant means entrusting data, sometimes personal data, to a service provider. Check it as you would any other online service, in line with the data-protection rules that apply to your company.

A data processing agreement

The assistant’s vendor processes data on your behalf. Check that a data processing agreement covers your plan and sets out what is processed, for how long and for what purpose, the types of data and each party’s obligations.

Sub-processors

Vendors rely on other providers to run their service. Read the list of sub-processors the vendor publishes, and check how you are told about changes.

Where the data is processed

Check where your data is processed and how transfers to other countries are governed. OpenAI, for example, offers data residency in Europe for its API, under conditions.

Source: OpenAI (Your data), accessed 4 October 2026. These pointers are general: they do not replace advice from your data protection officer or your legal adviser.

Health and social care: health data stays out of the connector by default

At Tech-O: a connector in production

The internal software C-Esium built for Tech-O, a French network of non-destructive water-leak detection companies, can be queried from Claude or ChatGPT through an MCP connector in production: about 35 tools, all read-only, grouped in these families.

Dashboard
Recent activity
Alerts
Today’s jobs
Unassigned jobs
Upcoming visits
Reports not done
Quotes and quote statistics
Invoices and unpaid invoices
Payments
Customers with unpaid invoices
Cash position
Revenue trend
History by address
Global search

Only the structure of the connector is shown here: no data from Tech-O or its customers is published.

Read the Tech-O case study

Ask for a demonstration

Frequently asked questions

What is MCP (Model Context Protocol)?

The Model Context Protocol is an open standard, published by Anthropic in November 2024, that connects AI assistants such as Claude or ChatGPT to external systems: databases, software, tools. C-Esium uses it so that your teams can query the data in their internal software from their assistant, read-only.

Can we query our company data from ChatGPT or Claude?

Yes, if your software exposes an MCP connector. Every internal software delivered by C-Esium includes one from V1: read-only, reserved for authorised users and logged, it lets you ask, for example, “Which quotes have been waiting for an answer for more than 15 days?” from Claude, ChatGPT or the app’s chat. To query your current tools without changing software, an AI connector is also available as a focused project, from AED 40,000.

Can the AI change or delete our data through the connector?

No. The MCP connector delivered by C-Esium is read-only: the assistant reads the data but cannot create, change or delete anything in the software. Access is reserved for the users you authorise and every access is logged.

Which ChatGPT or Claude plan should we choose for company data?

C-Esium recommends a business plan: ChatGPT Business or Enterprise, Claude Team or Enterprise, or the API. According to OpenAI and Anthropic, these plans do not use company data to train their models by default, and the administrator controls the connectors. The framework chosen is set out in the contract.

What about health data in health and social care?

It is excluded from the connector by default. Health data follows specific rules: C-Esium defines the hosting arrangement and the scope of the AI at scoping, in line with the applicable health-data rules, and by default limits the assistants’ access to data that is not health data (scheduling, HR, invoicing).

Sources

Official or primary sources, accessed 4 October 2026.