What is MCP?
MCP stands for Model Context Protocol. The specification describes it as an open protocol that enables seamless integration between LLM applications and external data sources and tools. In practice it is a shared language. An AI application, such as a chat assistant or a coding agent, uses it to ask a server two questions: what can you do, and please do this one thing with these inputs.
Anthropic introduced MCP on November 25, 2024 as an open standard. It is not a model, an agent framework, or a product. It is the wire format between them.
What problem does MCP solve?
An AI model is only as useful as the systems it can reach. Before MCP, each connection was built by hand. The announcement put it this way: “Every new data source requires its own custom implementation, making truly connected systems difficult to scale.”
Think of it from both sides. A company with an API had to write one integration for each AI client it wanted to appear in. A client builder had to write one connector for each service its users asked for. MCP turns that many-to-many job into one-to-many: the company publishes one server, the client implements one protocol, and the two meet.
How does MCP work?
The protocol uses JSON-RPC 2.0 messages between three roles: hosts (the AI application), clients (the connector inside the host) and servers (the service that provides context and capabilities). A typical exchange has four steps.
- The client connects and the two sides agree on a protocol version and capabilities.
- The client asks for the tool list. Each tool has a name, a description, and a JSON Schema for its inputs.
- The model reads those descriptions and decides to call a tool with specific arguments.
- The server runs the tool and returns a result the model can use in its answer or in the next call.
Messages travel over one of two standard transports: stdio, where the client launches the server as a local process, and Streamable HTTP, where the server is a URL that many clients can reach. A remote server can protect itself with the protocol's authorization flow, which is optional and based on OAuth 2.1.
The specification notes that MCP takes some inspiration from the Language Server Protocol, which did the same job for programming language support in code editors: one protocol, many editors, many languages.
Who has adopted MCP?
At launch, Anthropic named Block and Apollo as early adopters, and Zed, Replit, Codeium and Sourcegraph as development tool companies working with MCP.
A year later, in its December 9, 2025 post about donating MCP to the Agentic AI Foundation, Anthropic reported:
- more than 10,000 active public MCP servers;
- 97M+ monthly SDK downloads across Python and TypeScript;
- adoption by ChatGPT, Cursor, Gemini, Microsoft Copilot and Visual Studio Code, among other AI products.
The same post announced that MCP now belongs to the Agentic AI Foundation, a directed fund under the Linux Foundation that Anthropic, Block and OpenAI co-founded. For a company choosing what to build on, that matters: the protocol is governed in the open, not owned by one vendor.
How is MCP different from a plain API?
An API is written for a programmer who reads the documentation once and writes code against it. An MCP server is written for a model that reads the tool list at the start of every session and decides, on its own, what to call.
| REST API | MCP server | |
|---|---|---|
| Reader | A developer, before writing code | A model, at the start of each session |
| Discovery | Documentation or an OpenAPI file, read by a person | A tool list the client requests over the protocol |
| Integration cost | Custom code for each consumer | One server for every client that speaks MCP |
| Shape of an operation | Endpoint, verb, parameters | Named tool, plain description, JSON Schema input |
| Relationship | Does the work | Usually sits in front of the API and calls it |
Why should a company care?
Three reasons come up most often.
- Distribution. Your customers already work inside AI clients. A server puts your product where they are, without you building a plugin for each client.
- Control. You choose which operations agents may perform, how each is described, and who holds a key. That is safer than an agent improvising against your raw API.
- Visibility. When calls go through a server you run or host, you can see which tools are used, by which client, and which calls fail.
What does MCP not do?
MCP does not make a model smarter, and it does not decide what an agent is allowed to do. The specification is direct about this: it says tools represent arbitrary code execution and must be treated with appropriate caution, and that hosts must obtain explicit user consent before invoking any tool. Good tool descriptions, narrow permissions and call logs are still your job.
MCP also does not describe how two agents cooperate on a task. That is what A2A is for, covered in MCP vs A2A. And connecting many servers at once has a cost in context, covered in the too-many-tools problem.
Sources
Every figure and dated fact above comes from one of these pages. Follow the link to check it.
- Introducing the Model Context Protocol. Anthropic, November 25, 2024.
- Donating the Model Context Protocol and establishing the Agentic AI Foundation. Anthropic, December 9, 2025.
- Model Context Protocol specification, version 2025-06-18. modelcontextprotocol.io.
- MCP specification: Transports. modelcontextprotocol.io.
- MCP specification: Authorization. modelcontextprotocol.io.
