What is MCP, in one minute?
MCP, for Model Context Protocol, is an open protocol that standardises the connection between an AI application and outside tools. Anthropic published it in November 2024.
It works through 3 roles. The host is the AI application you use, Claude for instance. The MCP server exposes tools: find a client in the CRM, read a folder, retrieve a procedure. Between them, the MCP client passes the requests. Claude does not rummage through your files: it asks, and the server decides what it returns.
The protocol no longer belongs to a single vendor. Its governance sits with the Agentic AI Foundation, under the Linux Foundation. OpenAI adopted it in ChatGPT and its Agents SDK in April 2025, and Microsoft, Google and AWS support it. By mid-2026, more than 10,000 MCP servers were running in production according to published figures, and the official development kits were downloaded more than 97 million times a month.
What changed in 2026
A stateless protocol. On 28 July 2026, the MCP maintainers finalised the largest revision of the specification since authorisation was added. The protocol no longer tracks a session: each request carries what it needs. For a company, that means servers that are simpler to run behind a gateway, and a migration to plan for servers written in 2025.
Authorisation steered by the company. The Enterprise-Managed Authorization extension went stable this summer, adopted by Anthropic, Microsoft and Okta. The administrator manages access to MCP servers from the company identity provider, and the employee reaches all their servers with a single sign-in. The day they leave, access is cut in one place.
The flip side: booby-trapped servers
An MCP server has the rights you give it. If it reads your mail, it reads your mail. If it writes to your CRM, it writes to your CRM.
In September 2026, the Google Threat Intelligence Group reported booby-trapped copies of legitimate MCP servers published on PyPI, the public Python package repository. Same name, same description, different publisher. An employee who installs the wrong package thinking they are plugging in the official tool opens their data to a stranger.
It is the problem described in our article on shadow AI, one notch further: the employee no longer pastes a document into a chatbot, they give a piece of software permanent access to a folder.
The owner who bought Claude was right
We see it at our clients: owners equipped with Claude move fast and will not go back. They are right. The model is good, and it improves every quarter.
What Claude lacks inside your company is not intelligence. It is your documents, your rules and your limits: the current price list, the clauses you refuse to sign, the client who must not be chased again. MCP is the socket. The only question that matters is what is on the other side.
What we put on the other side
Forgeron3 exposes an MCP server. Claude plugs into it and finds your assistants, forged on your documents, and a reservoir of skills approved by the company: how a quote is priced here, how a tender is answered, what is and is not said to a client.
Forgeron3 does not replace Claude, it gives Claude what it does not know about you. The rules do not move when Claude plugs in: each assistant keeps a sealed scope, every access is logged and the log exports, your documents train no model, neither ours nor anyone else’s.
One point we would rather write down in black and white. When Claude queries the server, the passages it retrieves to answer enter the Claude conversation, at Anthropic. Your documents and their index stay hosted in France, the extracts used for the answer do not. If a file must never leave, the assistant stays in use inside Forgeron3, with no external connection. The setting is made assistant by assistant, and you are the one who makes it.
5 questions to ask before plugging in an MCP server
Whether the server comes from us, from a CRM vendor or from a public repository, the same questions apply.
If one of the 5 answers is « I do not know », the connection waits.
What is MCP (Model Context Protocol)?
An open protocol published by Anthropic in November 2024 that standardises the connection between an AI application and tools or data. The AI application calls an MCP server, and that server decides what it returns.
Is MCP an Anthropic product?
Anthropic created it, but the protocol is open and its governance sits with the Agentic AI Foundation, under the Linux Foundation. OpenAI, Microsoft, Google and AWS support it.
Can an MCP server change my data?
Yes, if its tools provide for it and you allow it. For a first connection, stay read-only, over a limited set of documents.
Does MCP replace an API?
No. An MCP server generally builds on your existing APIs and presents them in a format every compatible AI client understands. You write the access once, instead of one integration per AI tool.
Does my data stay in France if I use Claude with Forgeron3?
Your documents and their index, yes. The passages returned to Claude so that it can answer are processed by Anthropic. For sensitive files, the assistant stays in use with no external connection.
Sources
- Model Context Protocol, specification of 28 July 2026.
- Model Context Protocol, Enterprise-Managed Authorization extension promoted to stable, summer 2026, reported by InfoQ in July 2026.
- Google Threat Intelligence Group, « From prompting to autonomy: the evolution of adversarial AI », 8 September 2026.
- Server and SDK deployment figures: public data compiled in mid-2026.
Bring your Claude account and a folder of documents. We plug an assistant onto them in front of you: you see what Claude reads, what it cites, and what it does not see.
Request a demo→