The Model Context Protocol, or MCP, is a shared set of rules that lets an AI application connect to things outside itself: your files, a company database, a calendar, a search engine. Its own documentation compares it to a USB-C port. MCP aims to do the same for AI software and the tools it uses.

The problem MCP was built to solve

A large language model, the kind of system behind chatbots such as Claude and ChatGPT, only knows what was in its training data and what you type into it. To read your documents or act on your behalf, it needs a connection to other software.

Before MCP, each of those connections was built by hand. A developer who wanted an assistant to work with five services wrote five separate integrations, and another assistant needed five more. When Anthropic introduced MCP on 25 November 2024, it described the aim as replacing those one-off connections with a single open standard. The protocol was started by two Anthropic engineers, David Soria Parra and Justin Spahr-Summers. Early users named in the announcement included the payments company Block and the coding tools Zed, Replit and Sourcegraph.

How MCP works

MCP has three roles:

  • The host is the AI application you use, such as a chat assistant or a code editor.
  • The server is a small program that offers something useful, such as access to a database or a messaging service. It can run on your own computer or on the internet.
  • The client is the part of the host that holds the connection to one server. A host runs one client for each server it is connected to.

A server can offer three kinds of thing. Tools are actions the model can ask to run, such as sending a query to a database. Resources are data to read, such as the contents of a file. Prompts are reusable templates for common tasks.

In practice the exchange is simple. The host asks each server for a list of its tools. Each tool comes with a name, a plain-language description and the inputs it expects. The host passes that list to the model. When the model decides a tool would help, the host sends the request to the right server and feeds the result back into the conversation. The messages use JSON-RPC, a common format for one program to call another.

MCP does not make a model cleverer. It only standardises the wiring. The specification says the design borrows from the Language Server Protocol, an earlier standard that let code editors support many programming languages without custom work for each.

Who has adopted it

Adoption was quick. On 9 December 2025, a little over a year after launch, Anthropic gave MCP to the Agentic AI Foundation, a new body inside the Linux Foundation that it set up with Block and OpenAI. The Linux Foundation’s announcement lists Amazon Web Services, Bloomberg, Cloudflare, Google and Microsoft among the top-tier members, and names ChatGPT, Gemini, Microsoft Copilot, Cursor and VS Code among the products that support the protocol. Moving a standard to a neutral foundation is meant to reassure rivals that no single company controls it.

DateMilestoneFigure reported
25 November 2024Anthropic publishes MCPSix ready-made servers named, including Google Drive, Slack and GitHub
9 December 2025MCP donated to the Agentic AI FoundationMore than 10,000 active public servers; over 97 million monthly SDK downloads
28 July 2026Current specification releasedClose to half a billion monthly downloads across the main SDKs
MCP in numbers

An SDK, or software development kit, is the code library developers use to build MCP into their own products. These figures come from Anthropic and the MCP project, not from an independent count, and a download is not the same as a user. Automated build systems fetch the same library many times over.

Monthly downloads of the main MCP software kitsMillions of downloads a month
Monthly downloads of the main MCP software kits
ItemMillions of downloads a month
December 202597m+
July 2026Close to 500m

The two figures count slightly different sets of kits, and a download is not a user.Source: Anthropic and the MCP project. Chart: The Sidelight.

What changed in 2026

The protocol is still being revised. The version released on 28 July 2026 made MCP stateless, meaning every request carries all the information a server needs, with no opening handshake to remember. The same release tightened the rules for authorisation, the step where a server checks what a connected app is allowed to do, and marked several older features for removal, with at least twelve months’ notice.

Why security researchers are worried

The feature that makes MCP useful also makes it risky. A model reads tool descriptions and tool results as text, and it cannot reliably tell information from instructions. Hiding commands in that text is called prompt injection.

In April 2025 the security firm Invariant Labs described a “tool poisoning” attack. A malicious server offered a harmless-looking tool for adding two numbers. Its description, which the user did not see in full, told the model to read private files and send their contents back. The researchers demonstrated this against the code editor Cursor. They also described a “rug pull”, in which a server changes a tool’s description after the user has approved it.

A model reads tool descriptions and tool results as text, and it cannot reliably tell information from instructions.”

The specification is open about its limits. It says hosts must obtain a user’s consent before running any tool, and that tool descriptions from untrusted servers should not be believed. It then admits that MCP “cannot enforce these security principles at the protocol level”. The protection has to be built by whoever makes the app.

Companies that support MCP give similar warnings. OpenAI’s developer guide tells users that custom servers are third-party services it has not checked, and that a trusted developer’s server can still be unsafe if it handles untrusted content.

Governments have taken notice. On 20 May 2026 the US National Security Agency published design guidance on MCP. According to a summary by the law firm Reed Smith, the agency found that adoption had outpaced security safeguards. The risks it listed include access granted without identity checks and actions carried out without human sign-off. Its advice includes giving each tool the minimum access it needs and keeping detailed logs.

The project’s own security guidance adds a simpler risk. A server installed on your computer is ordinary software, and without restrictions it can run commands with the same privileges as the app that launched it.

What this means if you use AI tools

  • Treat a server like any other software you install. Use ones from publishers you recognise.
  • Read permission prompts. They are the main safeguard the standard relies on.
  • Limit what you connect. An assistant that can read your email and also browse the web can be tricked by one into leaking the other.

MCP has become the common way to connect AI models to the outside world in under two years, with backing from the largest AI companies. It has not settled how far to trust what comes back through the connection, and the sources reviewed here give no reliable figures on how often these attacks succeed outside the lab.