Drafted with Claude from official web pages (there is no paper) on 15 October 2026; reviewed by Zeeshan Arif. Read from anthropic.com (Anthropic announcement of 25 November 2024, read from the Web Archive copy of that day). Also read is the MCP specification, version 2024-11-05, at modelcontextprotocol.io (the overview and the transports page). MCP is a protocol, not a model, and has no research paper.
What it changed
It gave AI applications one open, shared way to connect to outside data and tools, so a tool written once as an “MCP server” can be used by any application that speaks the protocol, instead of each pairing needing its own custom connector.
The problem before it
Language models could already call tools, through function calling and agent loops like ReAct. But each application connected to each data source in its own way.
Anthropic’s announcement describes models as “constrained by their isolation from data”, trapped behind separate systems. Every new data source needed its own custom code, so connecting many applications to many sources did not scale.
The idea
The mechanism. MCP (Model Context Protocol) is an open standard for messages between AI applications and the systems that hold data and tools. Anthropic released it as open source with three parts: the specification with software development kits (SDKs), support for local MCP servers in the Claude Desktop apps, and an open-source collection of ready-made servers.
The specification names three roles:
- Hosts: AI applications, such as a chat app or a code editor, that start connections.
- Clients: connectors inside the host, one for each server.
- Servers: programs that offer data and abilities to the host.
A server can offer three kinds of things:
- Resources: data and context, for the user or the model to use.
- Prompts: ready-made message templates and workflows for users.
- Tools: functions that the model can run.
A client can offer servers sampling: a server may ask the host’s model to generate text, which allows agent-like behaviour that starts from the server’s side.
The messages use JSON-RPC 2.0, a standard format for one program to call another. Connections keep their state, and the two sides first agree which features each supports. The 2024-11-05 version defines two ways to carry the messages: standard input and output, for a server running on the same computer, and HTTP with Server-Sent Events, for a server elsewhere.
The specification says MCP takes inspiration from the Language Server Protocol, which lets any code editor support any programming language through one shared protocol.
The notation. None. It is a protocol, defined by a TypeScript schema.
What it showed
Reported by the sources. These are Anthropic’s own statements at launch, not measured results.
- Ready-made servers for Google Drive, Slack, GitHub, Git, Postgres and Puppeteer (a browser-control tool).
- Early adopters. Anthropic named Block and Apollo as having integrated MCP, and the development-tool companies Zed, Replit, Codeium and Sourcegraph as working with it.
- Building servers. Anthropic said Claude 3.5 Sonnet was good at writing MCP servers quickly.
- Rules for safety. The specification requires that users explicitly consent to data access and actions, that hosts get consent before sending user data to a server or running any tool, and that users approve any request from a server to use the model.
Limits, and what came later
The limits come from the sources:
- The protocol cannot enforce its own security principles. The specification says implementers “SHOULD” build consent flows, access controls and privacy protections, but MCP itself cannot require them.
- Tools mean running arbitrary code, and the specification warns that they must be treated with caution.
- At launch, Claude Desktop supported only local servers. Toolkits for remote servers that could serve a whole organisation were promised for later.
- The sources give no measured results, such as how many servers or users there were, or how reliable the connections were.
Later work depended on it. Claude Code used MCP to connect the coding agent to tools.
What to know first
- How language models call tools, as in ReAct.
- What a client and a server are, and what a protocol is: an agreed format for messages between programs.
- JSON, the text format the messages use.
How to read it
There is no paper to read. Instead:
- Read the announcement. It is short.
- Read the overview page of the specification, especially “Key Details” and “Security and Trust & Safety”.
- For more depth, read the specification’s pages on architecture and on server features (resources, prompts and tools).
Expect about 45 minutes for the steps above (an estimate).
Links