An AI assistant that can't reach your calendar, files, or tools can only talk about them. The Model Context Protocol (MCP) is the open standard for connecting AI applications to outside systems: data, tools, and services. Anthropic released it in November 2024, OpenAI adopted it in March 2025, Google announced support in April, and it now lives under the Linux Foundation.
The problem MCP solves
Before MCP, connecting AI to a tool was a custom job. Each AI application needed its own connector for each data source, so twenty apps talking to fifteen tools meant up to 300 separate integrations, each maintained by hand. Anthropic's announcement called even capable models "trapped behind information silos," with every new data source requiring its own custom implementation.
MCP replaces the custom wiring with one standard. A tool writes its connection once, as an MCP server. An AI application adds MCP support once, as a client. A compatible client can then reach a compatible server without a bespoke connector for that pair, subject to what the host supports and the permissions you grant. When a company publishes an MCP server for its product, supported AI apps may be able to use it without building their own connector.
How it works, in plain language
MCP has two halves. On the tool side, an MCP server is a small program that exposes one data source or service in the standard format: your Google Drive, a Slack workspace, a GitHub repo, a Postgres database. On the AI side, the client lives inside the host application (Claude Desktop, ChatGPT, an agent tool) and connects to the servers you approve.
A server can expose information to the AI, and when it provides tools, the AI can request actions through it: create a document, file an issue, run a search. You and the host app still control what's allowed. Connect a Drive server and a GitHub server to the same assistant, for example, and one request can pull last quarter's brief from Drive and file a summary as an issue.
The protocol launched with pre-built servers for common systems (Google Drive, Slack, GitHub, Git, Postgres), and companies publish servers for their own products, which is how the library keeps growing.
Where you'll meet it
Claude Desktop installs local MCP servers through desktop extensions: pick one from the directory, click install, and it runs on your computer, no code involved.
ChatGPT added MCP support starting in March 2025, beginning with the Agents SDK and the desktop app.
Agent and coding tools lean on MCP for their reach. The protocol's documentation lists an agent reading your calendar, a coding tool building a web app from a Figma design, and AI creating 3D designs in Blender, then printing them on a 3D printer.
Who owns it now
Since December 9, 2025, MCP has belonged to a foundation rather than a company. Anthropic created it and ran it for the first year, then donated it to the Agentic AI Foundation, a directed fund under the Linux Foundation, co-founded by Anthropic, Block, and OpenAI, with Google, Microsoft, AWS, and Cloudflare among the supporters. For a standard meant to be shared, governance matters: the protocol's direction is no longer one vendor's call.
What MCP doesn't do
MCP gives an AI access; it doesn't decide whether an action is wise. The results still need review. How to Verify AI Output: A Simple Framework covers that part.
Grant access narrowly as well. A server may act with the permissions you grant it, so give it the smallest slice that does the job, and read what a connector asks for before you install it.
Setup has limits too. Ready-made servers install in a click; a custom server for your own systems takes developer work. And MCP doesn't make an AI an agent: what the AI does once it's connected is a different question.
FAQ
What does MCP stand for?
Model Context Protocol. The name describes the job: a protocol (a shared set of rules) for providing context (your data, tools, and services) to AI models.
Is MCP the same as an API?
An API is one service's custom door: it connects that service to whatever the developer builds. MCP is a shared rulebook for how AI applications talk to many such doors, so one AI app can reach many services without custom wiring for each. MCP servers themselves often use APIs under the hood.
Do I need to know how to code to use MCP?
Not for ready-made servers. Claude Desktop's extension directory installs them in one click, no code involved. Building a custom server for a system nobody has published yet takes developer work.
Is MCP owned by Anthropic?
Anthropic created MCP and donated it to the Agentic AI Foundation under the Linux Foundation on December 9, 2025. The foundation is co-founded by Anthropic, Block, and OpenAI, so changes to the protocol go through a shared process rather than one company's product roadmap.
Which AI tools support MCP?
Claude's apps, OpenAI's products (from March 2025), and Google's Gemini models and SDK (announced April 2025) all support it. Support changes, so check your tool's documentation for what's current.
What's the difference between MCP and an agent?
An agent is a program that takes multi-step action toward a goal with limited supervision. MCP is the connection standard that can hand that agent its tools. Many agent tools use MCP; the two answer different questions.
Want to go further?
- What Are AI Agents in 2026? shows where MCP fits in the agent stack.
- What Is Agentic AI? The Next Step Beyond Chatbots defines the behavior these connections enable.