What is MCP? Connecting company data to Claude or ChatGPT
MCP (Model Context Protocol) is an open-source standard for connecting AI applications such as Claude or ChatGPT to external systems: data sources, tools and reusable prompts. A company writes one MCP server in front of its data, and any application that supports MCP can use it. The protocol carries the exchange; what the model does with the data, and who is allowed to see what, still has to be designed.
If you use Claude or ChatGPT at work, you have probably pasted a document into the window to get an answer about it. MCP is the standard that lets the assistant fetch the document itself, from the place where it already lives.
What is MCP?
The official documentation opens with this definition: MCP (Model Context Protocol) is an open-source standard for connecting AI applications to external systems. Through MCP, applications like Claude or ChatGPT can connect to data sources such as local files and databases, to tools such as search engines and calculators, and to workflows such as specialised prompts. That gives them access to key information and lets them perform tasks.
The documentation compares MCP to a USB-C port for AI applications. A USB-C port gives electronic devices one standard way to connect; MCP gives AI applications one standard way to connect to external systems. The documentation’s promise to builders is to build once and integrate everywhere.
How does MCP work?
MCP follows a client-server architecture with three participants.
- The host is the AI application, for example Claude Desktop or Claude Code.
- The client is a component inside the host. The host creates one client for each server, and each client keeps a dedicated connection to its server.
- The server is a program that provides context to the clients.
A server can run on your machine or elsewhere. The documentation calls a server that runs locally and talks through standard input and output a “local” server, and one hosted on a platform and reached over HTTP a “remote” server. Remote servers use the Streamable HTTP transport, which supports bearer tokens, API keys and custom headers; MCP recommends OAuth to obtain the tokens.
The messages themselves use JSON-RPC 2.0. The documentation adds a precise limit: MCP focuses on the protocol for exchanging context, and it does not dictate how AI applications use language models or manage the context they receive.
What can an MCP server expose?
A server offers three kinds of building blocks, and the documentation says who controls each.
| Building block | What it is | Who controls it |
|---|---|---|
| Tools | Functions the model can call, such as searching flights, sending a message or creating a calendar event | The model decides when to use them |
| Resources | Read-only data for context, such as file contents, database schemas or documentation | The application |
| Prompts | Ready-made instruction templates, such as “summarize my meetings” | The user, by explicit invocation |
Tools can write to databases, call external APIs and modify files, so the documentation stresses human oversight. Tools may require user consent before they run. Applications can show the available tools, ask for approval on each execution, let users pre-approve safe operations and keep activity logs of every call.
Why connect company data to Claude or ChatGPT?
The documentation lists what MCP can enable, among them enterprise chatbots that connect to several databases across an organisation so that users can analyse data by chatting. It states the benefit for each side: less development time for builders, access to an ecosystem of data and tools for applications, and more capable assistants for end users, who can reach their own data and have actions taken on their behalf when necessary.
In a company, the practical point is that your documents and your business software stay where they are. You expose them once through a server, and the assistants your staff already use can query them. Several applications support the standard, so the same server can serve more than one of them.
The limit is the one quoted above. MCP carries the data to the model. Whether the answer is correct depends on how well the data is organised and on the tools you expose: an assistant given a search tool that returns the wrong files will give wrong answers with perfect politeness.
What are the security risks of MCP?
The official security page is specific. A few of the risks it describes:
- Confused deputy: a proxy server that fronts a third-party API can be tricked into handing an authorisation code to an attacker. The page requires MCP proxy servers to obtain consent for each client.
- Token passthrough: a server that accepts tokens not issued to it and forwards them downstream. The page states that MCP servers must not accept tokens that were not explicitly issued for them.
- Local server compromise: a local server runs with the same privileges as the client, so a malicious startup command or package can execute code on the machine. Clients must show the exact command and ask for approval.
- Scope minimisation: broad permissions granted up front increase the damage if a token leaks. The page recommends least-privilege scopes.
So the questions to ask before connecting company data are concrete. Who built the server, where does it run, which permissions does it request, and who sees which data.
And with us?
We publish this blog. We are a forward deployed agency: three engineers from Télécom Paris, specialised in AI, who build custom AI agents and also take on machine learning and data missions. Our agents cite their sources, write Word and PowerPoint files, and are reachable from Claude, ChatGPT or Copilot as well.
On the data side, sources are read in read-only mode from SharePoint, OneDrive or Dropbox, in folders the client picks, and no file is kept; other tools can be connected on request. Each company has its own encryption key, servers are in France and the database in the European Union, the models are European (Mistral AI) under a contract with no training on the data, and everything can be erased on request. Nothing goes out without validation by a person at the company. More on this on the page on data security, and on connecting Claude or ChatGPT to company data.
Sources, consulted on 12 November 2026
Questions
What does MCP stand for?
Model Context Protocol. The official documentation defines it as an open-source standard for connecting AI applications to external systems.
Is MCP only for Claude?
No. The documentation says MCP is supported by AI assistants such as Claude and ChatGPT, and by development tools such as Visual Studio Code and Cursor.
Does MCP make my data safe?
MCP defines how an application and a server talk to each other. Safety depends on how the server is built and what it is allowed to do. The official security page lists risks such as confused deputy attacks, token passthrough and compromised local servers, and the rules servers must follow against them.
Do I need a developer to use MCP?
Using a server that already exists means connecting it in your AI application. Exposing your own documents and databases means someone has to build and secure a server, which is engineering work.