Definition
Model Context Protocol (MCP) is an open protocol that standardizes how AI applications connect to external tools, data sources and services. Anthropic open-sourced it in November 2024, and since December 9, 2025 it has been a founding project of the Agentic AI Foundation (AAIF), a directed fund under the Linux Foundation.
Before MCP, every pairing of an AI assistant and a business system needed its own integration. MCP replaces that with one interface: a developer builds an MCP server once, and any MCP-compatible client can use it. The protocol borrows its model from the Language Server Protocol, which did the same for programming languages and code editors.
The architecture has three roles. A host is the AI application the user works in. A client is the connector inside the host. A server exposes capabilities, which come in three types: tools (functions the model can call, similar to function calling), resources (data the model or user can read) and prompts (reusable templates). Messages use JSON-RPC 2.0 over a local stdio connection or Streamable HTTP. Optional extensions add asynchronous Tasks, Skills over MCP and MCP Apps, which render interactive UI inside a conversation.
The current specification revision is dated 2026-07-28 and replaced 2025-11-25. It made the protocol stateless by removing the initialize handshake and protocol-level sessions, added a server/discover method for version negotiation, introduced multi round-trip requests for collecting user input, and deprecated the Roots, Sampling and Logging features. It also adopted a feature lifecycle policy that guarantees at least 12 months between deprecation and removal, and tightened OAuth by requiring clients to validate the issuer in authorization responses. The full list is in the 2026-07-28 changelog.
Governance sits with the project's maintainers, core maintainers and lead maintainers, and specification changes go through Specification Enhancement Proposals (SEPs). The AAIF governing board handles budget, membership and new projects but, as the MCP team's announcement put it, does not dictate technical direction. Anthropic, Block and OpenAI co-founded the foundation with support from Google, Microsoft, AWS, Cloudflare and Bloomberg, and it reported 170 member organizations at MCP Dev Summit North America in 2026. At the same event, co-creator David Soria Parra said MCP SDKs were downloaded more than 110 million times a month, with OpenAI's Agents SDK and LangChain both pulling MCP in as a dependency.
For search and marketing teams, MCP is not a discovery channel in the way crawling is. ChatGPT, Google and Perplexity still find and cite public pages through their crawlers and indexes, and MCP plays no part in that. Its relevance shows up in three other places. First, AI agents that act for users (comparing products, booking, filing tickets) increasingly reach a company's catalog, docs or account data through an MCP server, which makes server quality part of agent readiness alongside protocols such as Agent2Agent and the Agentic Commerce Protocol. Second, developer products publish docs and API servers so coding agents can use them correctly. Third, analysts connect data tools to Claude or ChatGPT through MCP; Promptwatch, for example, exposes its visibility data this way, and our post on using the Promptwatch MCP in Claude shows the workflows.
Security needs attention. Tool descriptions and returned content are untrusted input that can carry prompt injection, and researchers demonstrated DNS rebinding attacks against local MCP servers in 2026. The spec requires explicit user consent before a host invokes a tool. For how MCP relates to agent payment and checkout protocols, see our comparison of AP2, ACP and MCP.
Examples of Model Context Protocol (MCP)
- Uber engineers described at MCP Dev Summit North America 2026 how an internal MCP gateway turns service endpoints into tools for more than 1,500 monthly active agents, with third-party servers held to a stricter review than internal ones.
- A B2B software vendor publishes a read-only MCP server for its public docs and pricing. When a prospect's agent compares vendors, it queries current plan limits directly instead of relying on a third-party review page from last year.
- An agency connects an AI visibility platform's MCP server to Claude and asks for every prompt where a client lost citations to a competitor this month, then has Claude draft the client report from the returned data.
- A team running a local MCP server on developer laptops enables origin header validation after learning that a malicious web page could otherwise reach the server through DNS rebinding.
