Model Context Protocol

In short

The Model Context Protocol (MCP) is an open protocol for connecting language-model applications to external data sources and tools². Anthropic released it so developers could build against one standard instead of a separate connector for each data source¹.

Also called: MCP
Topic: AI agents
Level: Some background helps
First described: Anthropic introduced and open-sourced MCP on 25 November 2024. Its announcement names David Soria Parra and Justin Spahr-Summers as its creators¹.

MCP uses JSON-RPC 2.0 messages between three roles. A host is the language-model application that starts a connection. A client is a connector inside that host. A server provides context and capabilities². A server can offer resources, which are context and data for the user or the model; prompts, which are templated messages and workflows; and tools, which are functions for the model to execute². The base protocol includes server and client capability negotiation².

Anthropic’s announcement describes two sides of that design. Developers either expose data through MCP servers, or they build AI applications, called MCP clients, that connect to those servers¹. The stated aim is to replace separate connectors, one per data source, with a single protocol¹. At the launch, Anthropic shared pre-built servers for Google Drive, Slack, GitHub, Git, Postgres and Puppeteer¹.

The specification says MCP itself cannot enforce these security principles at the protocol level. Users must explicitly consent to and understand all data access and operations, and a host must obtain explicit user consent before it invokes a tool². Tools represent arbitrary code execution. Descriptions of tool behaviour, such as annotations, should be treated as untrusted unless they come from a trusted server². The tools section adds that there should always be a person who can deny a tool call³.

Example

The tools specification uses a weather lookup, get_weather, as its sample tool. The client lists the server’s tools, the model selects one, and the client sends the call³. Separately, the specification says the host must get explicit user consent before invoking any tool².

Why it matters

MCP is a shared way for a language-model application to reach external data and tools, including tools that run code, and the host must get explicit consent before a tool call². The tools section also says a person should be able to deny that call³.

Questions people ask

What does MCP stand for?

MCP stands for Model Context Protocol. The specification calls it an open protocol for connecting language-model applications with external data sources and tools².

Is MCP the same as an API?

No. The specification defines MCP as a protocol that uses JSON-RPC 2.0 messages between a host, a client and a server². A tool may query a database, call an API, or run a computation³. Anthropic describes MCP as a replacement for a separate connector for each data source¹.

Can an MCP tool take an action, or only supply data?

Both exist, as different features. Resources supply context and data. Tools are functions the model can execute, and the specification treats them as arbitrary code execution². A tool can query a database, call an API, or run a computation³. The host must obtain explicit user consent before it invokes a tool².

Related terms

Tools that use it

Sources

Back to the glossary