The Model Context Protocol (MCP) is an open standard that lets AI apps connect to tools and data through one shared interface. Instead of a custom integration for every pairing, an app that speaks MCP can use any MCP server. Anthropic introduced it on November 25, 2024.
This page is the reference definition: what the standard is, who maintains it and how its spec versions work. For a beginner walkthrough with examples, read What is MCP?.
Definition in one paragraph
MCP is a specification, not a product. It defines how a host application and a server exchange messages so a model can discover and use capabilities at runtime. The wire format is JSON-RPC 2.0. A server can offer three kinds of capability: tools the model can call, resources the app can read, and prompts a user can pick. Anyone can implement either side, and official SDKs exist for several languages.
Roles and transports
MCP has three roles:
- The host is the AI app you use, like Claude Desktop or an IDE.
- The client is the connector inside the host. It holds one connection to one server. See MCP client.
- The server is the program that offers capabilities. See MCP server.
Messages move over one of two transports: stdio for a server on your own machine, or Streamable HTTP for a server across the network, which is what makes a server remote.
Who maintains MCP
MCP was designed at Anthropic by David Soria Parra and Justin Spahr-Summers. Block and Apollo were early adopters, along with dev-tool companies like Zed, Replit, Codeium and Sourcegraph.
As of October 2026, MCP is stewarded by the Agentic AI Foundation (AAIF), a fund under the Linux Foundation. Anthropic donated the project to it on December 9, 2025. Anthropic, Block and OpenAI co-founded it, with support from Google, Microsoft, AWS, Cloudflare and Bloomberg.
Day-to-day technical decisions stay with MCP's own maintainers. Changes to the spec go through public proposals called SEPs (Specification Enhancement Proposals), reviewed in the open rather than set by one company's roadmap. The reference MCP Registry is a separate community project and is still labeled a preview.
Spec versions
MCP versions its specification by release date, written as YYYY-MM-DD, not by semantic version number. Revisions named in this glossary include 2024-11-05, 2025-03-26, 2025-06-18, 2025-11-25 and 2026-07-28. Clients and servers agree on a revision when they connect, and the spec defines backward-compatibility rules for older ones.
The biggest shifts so far touched the transport, authorization and how tools describe themselves. The 2026-07-28 revision changed Streamable HTTP again. MCPifex does not claim support for that revision. The full history is in MCP spec changes.
What MCP is not
MCP is not an agent framework, a model or a hosting service. It does not decide which tool to call; the model does. It does not store your data. It only standardizes the connection. Related protocols such as A2A cover agent-to-agent traffic, and the MCP vs A2A comparison explains where each fits. Security is still the operator's job: see the MCP security checklist.
Why governance matters
A protocol that one vendor controls is a risk for everyone building on it. Moving MCP under a neutral foundation, with public proposals and several maintainers, lowers that risk. It also explains why clients from competing companies, including Claude, ChatGPT and Cursor, can all connect to the same servers.
Where MCPifex fits
Running servers yourself is the hard part. You can find them through a registry or use a hosted gateway. MCPifex hosts servers on the same spec: pick one on the marketplace, and any MCP client that supports Streamable HTTP can connect to https://mcp.mcpifex.com/mcp. We run the server for you.
Quick answers
Anthropic created MCP and donated it to the Agentic AI Foundation, a fund under the Linux Foundation, on December 9, 2025. MCP's own maintainers make day-to-day technical decisions, and spec changes go through public SEPs.
By release date in YYYY-MM-DD form, such as 2025-06-18 or 2025-11-25. Client and server agree on a revision when they connect, and the spec defines backward-compatibility rules.
No. An API is a fixed interface for one service. MCP is a shared protocol that lets a model discover and call many servers' tools at runtime. See the MCP vs API comparison for details.
Sources
- Introducing the Model Context Protocol — Anthropic
- Governance and Stewardship — Model Context Protocol
- MCP joins the Agentic AI Foundation — Model Context Protocol Blog
- Versioning — Model Context Protocol Specification
Ready to try it?
Run hosted MCP servers behind one endpoint and choose exactly which tools your AI can use.