Last updated: 2026-11-20 (protocol revision 2026-07-28)
The short answer
The Model Context Protocol (MCP) is an open standard for connecting AI applications to external tools and data. An MCP server wraps something useful, such as a database, a file system, or a SaaS API, and describes it in a common format. Any MCP-compatible AI application can then use it without custom integration code.
Anthropic open-sourced MCP on 25 November 2024 and donated it in December 2025 to the Agentic AI Foundation, a fund under the Linux Foundation. The specification lives at modelcontextprotocol.io.
The problem it solves
Before MCP, every AI application needed its own connector for every service. Ten applications and ten services meant up to a hundred separate integrations, each with its own auth handling, schema, and bugs.
MCP turns that into a shared interface. A service publishes one MCP server. An application implements one MCP client. Any client can talk to any server.
How MCP works
MCP has three roles:
| Role | What it is | Example |
|---|---|---|
| Host | The AI application the person uses. It decides which servers to connect to and asks for consent. | A chat app, an IDE, an agent workspace |
| Client | A connector inside the host. Each client talks to one server. | Created and managed by the host |
| Server | A program that exposes tools and data in the MCP format. | A GitHub server, a Postgres server |
Messages are JSON-RPC. There are two standard transports:
- stdio: the host starts the server as a local process and talks to it over standard input and output. Common for servers that touch local files.
- Streamable HTTP: the server runs remotely and the client reaches it over HTTP. Common for hosted services.
What a server can expose
A server offers up to three kinds of things:
- Tools: functions the model can call, such as
search_issuesorrun_query. Each tool has a name, a description, and an input schema. - Resources: data the application can read, such as a file, a database row, or a document.
- Prompts: reusable prompt templates that the person can pick, such as "review this pull request".
A server can also ask the person for a piece of input through the client. This is called elicitation.
Two older client features are being retired. Sampling, where a server asked the client for a model completion, and roots, where the client told a server which folders it could use, are deprecated as of the 2026-07-28 revision. New servers should not rely on them.
What changed in the current revision
MCP versions are dates. The current revision is 2026-07-28, and it made the protocol stateless:
- There is no connection handshake. Every request carries its own protocol version and client capabilities.
- Servers advertise what they support through a
server/discovercall. - Protocol-level sessions were removed from the HTTP transport.
- Servers no longer send requests to clients. When a server needs something, such as an answer from the person, it replies that input is required and the client retries the request with the answer.
- Sampling, roots, and protocol-level logging are deprecated. They stay in the specification for at least twelve months and can be removed in a revision released on or after 28 July 2027.
For most people this is invisible. It matters if you build or host servers, because a stateless server is simpler to scale and to put behind a load balancer. Older clients and servers written for earlier revisions may need an update to talk to newer ones.
MCP and function calling
Function calling is how a model asks its application to run a function. MCP is how the application finds and reaches those functions in the first place.
They work together. The host lists the tools from its MCP servers, passes them to the model as callable functions, and when the model calls one, the host forwards the call to the right server.
Is MCP safe?
MCP is a transport for capabilities, so the risk depends on the server. A server that can read your files or send email can do those things when the model asks.
Check four things before connecting a server:
- Who publishes it. An official server from the service itself carries less risk than an anonymous fork.
- What tools it exposes. Read the tool list. A read-only server is a smaller risk than one that can write or delete.
- How it authenticates. Prefer per-user sign-in over a shared long-lived key.
- Where it runs. A local stdio server has the permissions of your user account. A remote server sees whatever you send it.
How to start using MCP
- Pick a host that supports MCP. Most major AI assistants and coding tools do.
- Find a server for the service you want. Directories such as the SourceWeft MCP directory list servers with their tools, authentication method, and connect guides.
- Add the server to your host. A remote server needs a URL and a sign-in. A local server needs a command to run.
- Ask the assistant to do something that needs the server, and approve the tool call when asked.
FAQ
Who created MCP? Anthropic open-sourced it on 25 November 2024. In December 2025 Anthropic donated it to the Agentic AI Foundation, which is part of the Linux Foundation. The specification and SDKs are open source.
Is MCP only for Claude? No. MCP is model-agnostic. Any application can implement a client, and many assistants, IDEs, and agent frameworks have.
What is an MCP server? A program that exposes tools, resources, or prompts in the MCP format, so that AI applications can use a service without a custom integration.
What is the difference between MCP and an API? An API is designed for developers to call from code. An MCP server describes its capabilities in a way a model can discover and call at run time. Many MCP servers are thin wrappers over an existing API.
What is the difference between MCP servers and agent skills? An MCP server gives an agent new capabilities by connecting it to a system. A skill is a folder of instructions and files that teaches an agent how to do a task well. See MCP servers vs agent skills.
Do I need to code to use MCP? No. Using a server is configuration: add a URL or a command to your host. Building a server needs code, and official SDKs exist for the common languages.