The Model Context Protocol, or MCP, is an open messaging standard that lets an AI application connect to outside data and tools through one shared interface instead of a separate integration for every service. Anthropic released it as open source on November 25, 2024, together with developer SDKs, local server support in Claude Desktop, and a library of prebuilt connectors. Two years later it is maintained outside Anthropic, under a dedicated open-source foundation, and its own specification has already gone through several revisions.
What problem it solves
Before MCP, an AI application that wanted to read a calendar, query a database or call a code search tool needed its own custom connector for each one. The protocol specification compares the approach to the Language Server Protocol, which let one editor support many programming languages through a single interface instead of writing a separate plugin for each language and tool. MCP does the same job for AI applications and the external systems they need context from.
The specification fixes three roles. A host is the AI application that starts a connection, such as a desktop assistant or an IDE plugin. A client is the connector inside that host. A server is the program that actually provides context and capabilities, such as a wrapper around a file system, a ticketing system or a search index. Messages between client and server use JSON-RPC 2.0, the same remote-procedure-call format used by several existing developer tools.
What a server can offer
An MCP server can expose three kinds of capability to a host:
| Capability | What it gives the host |
|---|---|
| Resources | Data the user or the model can read, such as a file, a database record or a log |
| Prompts | Templated messages or multi-step workflows the host can offer the user |
| Tools | Functions the model can call and get a result back from |
A client and server negotiate which of these a given connection actually supports before doing any work, so a minimal server that only exposes one tool is a valid MCP server.
What changed for Sampling, Roots and Logging
Early MCP specifications also defined client-side capabilities a server could call on: sampling let a server ask the client to run a language-model completion on its behalf, roots let a server ask which file or URI boundaries it was allowed to operate within, and logging let a server set its log verbosity through the protocol itself. The project’s 2026-07-28 specification changelog deprecates all three, alongside the initialize handshake and the Mcp-Session-Id header that earlier versions used to open a stateful connection; every request now carries its own protocol version and capabilities instead.
| Capability | Status in the 2025-06-18 and 2025-11-25 specs | Status in the 2026-07-28 spec |
|---|---|---|
| Sampling | Current client capability | Deprecated; suggested replacement is calling an LLM provider’s API directly |
| Roots | Current client capability | Deprecated; suggested replacement is a tool parameter or static server configuration |
| Logging | logging/setLevel set the server’s log level over the protocol | Deprecated; ping and logging/setLevel are removed, with stderr or OpenTelemetry suggested instead |
| Elicitation | Current; lets a server ask the user a question through the host | Still current, with changes to how multi-choice answers and default values are expressed |
The changelog notes that deprecated features “remain fully functional during the deprecation window,” which the project’s governance rules set at a minimum of twelve months, so an older server built around sampling or roots keeps working rather than breaking outright.
Is MCP tied to one AI company or model?
No. Anthropic designed and published the original specification, but on December 9, 2025 it donated MCP to the Agentic AI Foundation, a directed fund the Linux Foundation now runs. Anthropic, Block and OpenAI co-founded that foundation, with Google, Microsoft, Amazon Web Services, Cloudflare and Bloomberg listed as supporters; Anthropic’s own announcement says MCP’s governance model “will remain unchanged” under the move. Nothing in the specification requires a particular model or vendor on either side of a connection: any application that can hold up its end of the JSON-RPC exchange can act as a host, and any program that exposes resources, prompts or tools in the expected format can act as a server.





