What is MCP? The Model Context Protocol, explained
An open standard connecting AI apps to tools, data, and prompts: a USB-C port for AI, in its own docs' words. Where it came from, how it works, and how it differs from an API.
MCP, the Model Context Protocol, is an open standard for connecting AI applications to external systems: the tools they can call, the data they can read, the prompt templates they reuse. The project’s own analogy is a USB-C port for AI applications: one standardized connector in place of a custom integration per app, per tool. Anthropic created it in November 2024 and open-sourced it; since December 2025 it lives under the Linux Foundation, and it’s spoken today by Claude, ChatGPT, Gemini’s API, and open-source agents like OpenClaw.
The problem it exists to solve
A model on its own can only talk. To do anything (read your files, search the web, query a database, file a ticket), the application around it has to wire up each capability. Before MCP that wiring was bespoke: every AI app built its own connector for every tool, an every-app-times-every-tool matrix that mostly went unbuilt. MCP replaces the matrix with a protocol. An AI application (the host) runs a small client for each connection, and any number of servers each expose three kinds of thing: tools (functions the model can invoke), resources (data it can read), and prompts (reusable templates). Build a server once and every MCP-speaking app can use it; build a host once and it can reach every MCP server in the ecosystem.
How it works, one level down
Messages are JSON-RPC 2.0 (the spec requires it) over one of two transports: stdio, where the client launches the server as a local subprocess and talks over its standard streams, or Streamable HTTP, where each message is an HTTP POST to a remote endpoint. As of the current spec revision (2026-07-28), MCP is a stateless protocol: every request carries everything needed to process it, which means a request can land on any server instance behind an ordinary load balancer, the change that made remote MCP servers easy to scale. Heavier machinery (long-running tasks, interactive apps, enterprise-managed authorization) lives in official extensions rather than the core, which stays deliberately small.
Is MCP an API, then?
Close, but the distinction matters. An API is one service’s own interface (Stripe’s API, GitHub’s API), and each one is different by design. MCP standardizes the model-facing side: the AI application speaks one protocol and gets tool discovery, invocation, and context in one shape, whatever sits behind it. In practice most MCP servers are thin adapters that wrap an ordinary API and present it to models in the standard form. So MCP doesn’t replace APIs: it’s the layer that stops every AI app from having to learn each of them separately.
Who’s behind it now
MCP was created at Anthropic by David Soria Parra and Justin Spahr-Summers, and Anthropic donated it to the Linux Foundation’s new Agentic AI Foundation in December 2025, a fund co-founded by Anthropic, Block, and OpenAI, with support from Google, Microsoft, AWS, and Cloudflare. That breadth shows in the tooling: OpenAI’s developer docs call MCP “an open protocol that’s becoming the industry standard for extending AI models” and document it across ChatGPT and Codex, and Google’s Gemini API supports remote MCP servers directly. By Anthropic’s count at donation time, more than 10,000 public MCP servers existed, with SDK downloads running at 97 million a month.
MCP and the other protocol next to it
You’ll often see MCP named beside ACP, the Agent Client Protocol. They’re complements: MCP connects an agent to its tools and context; ACP connects an agent to the editor or UI a person drives it from. A single agent process, a Hermes or an OpenClaw, can speak both at once, and OpenClaw in particular sits on both sides of MCP: it can consume MCP servers and expose its own conversations as one. What that looks like in practice, and how to run servers without handing your machine to a stranger, is the subject of the MCP server guide.