Model Context Protocol (MCP)

The Model Context Protocol, usually shortened to MCP, is an open standard for connecting AI assistants to outside systems. A product exposes its data and its actions as an MCP server. An assistant connects to that server as a client, discovers what it offers, and calls it during a conversation. One server works with every client that speaks the protocol, so a product no longer builds a separate integration for each assistant.

The name describes the job. Context is what a model needs to know and does not have in the conversation. Protocol is the shared grammar for going and getting it. Before MCP, giving an assistant access to a calendar, a database or a file store meant building to that assistant's own plugin format and then rebuilding for the next one. MCP replaced a matrix of integrations with one interface on each side.

A connection has three moving parts. The server advertises a list of tools, each with a name, a description and a typed schema for its arguments. The client puts those in front of the model, which picks one and fills in the arguments. The server runs it and returns a result the model reads as part of the turn. Transport is JSON-RPC, over stdio for a server running on your own machine and over HTTP for a remote one. Remote servers authorize with standard OAuth, so a connection is granted and revoked like any other app you have signed in to. Servers can also expose resources, which are readable documents, and prompts, which are named instructions a user can invoke directly.

MCP is governed by a foundation rather than by a single vendor. Anthropic published it in late 2024 and donated it to the Agentic AI Foundation, a Linux Foundation directed fund co-founded with Block and OpenAI and backed by Google, Microsoft, AWS and Cloudflare. By early 2026 the ecosystem had passed 10,000 active servers and roughly 97 million monthly SDK downloads, with first-class client support in Claude, ChatGPT, Cursor, Gemini, Copilot and VS Code. The precedent people reach for is Kubernetes going to the CNCF in 2016: neutral stewardship is what made the standard safe to build on.

For a user the practical effect is that the assistant you already pay for gains abilities it did not ship with, and those abilities follow you when you change assistants. For a product it means the assistant stops competing with your interface. It becomes your interface.

Two limits are worth knowing before you connect anything. First, a tool the model cannot see is a tool it concludes does not exist, so a server exposing fifty tools spends a large share of the context window on their schemas, and a server that hides them behind a discovery step has to make discovery good. Second, a tool result is text arriving inside the conversation, so a page or an email fetched by a server can carry instructions aimed at the model rather than at you.

That second limit leads to the question MCP deliberately does not answer: what a write is allowed to do. A tool call is a function call, so a conformant server may commit whatever the model asks for, or may hold it for a person to approve. Both pass the specification. The difference is invisible in a tool listing and decisive once the data describes real people.