What is MCP (Model Context Protocol)?

What MCP is

Anthropic's announcement on 25 November 2024 described MCP as "an open standard that enables developers to build secure, two-way connections" between data sources and AI-powered tools, replacing a separate integration for every pair. The specification, maintained at modelcontextprotocol.io (its current schema is dated 28 July 2026), defines three roles: hosts are the AI apps, clients are the connectors inside them, and servers provide context and capabilities. Messages use JSON-RPC 2.0.

A server can offer three kinds of things. Tools are functions the model can run, such as "search tickets" or "create invoice". Resources are data for the user or the model to read. Prompts are templated messages and workflows. The specification also lists optional extensions, including one for serving skills: rich instructions for agent workflows, discovered and used through MCP.

On 9 December 2025, Anthropic donated MCP to the new Agentic AI Foundation, a directed fund under the Linux Foundation co-founded with Block and OpenAI, where it joined Block's goose and OpenAI's AGENTS.md as founding projects. Anthropic says MCP's governance model stays the same, with its maintainers leading.

Why a small team should care

• Your AI can work in your tools. With a connector, the assistant reads the help center, files the ticket or checks the calendar instead of you pasting things in.
• One connection, many AIs. The same server works in Claude, ChatGPT and coding agents, so a teammate on a different AI gets the same tools.
• It is also a risk surface. The specification is blunt: tools represent arbitrary code execution, a host must get the user's consent before invoking one, and tool descriptions should be treated as untrusted unless the server is trusted.

A worked example

Illustration: a made-up business. Pinegrove Studio, an eight-person design agency, wants its AI to read the project tracker when writing client updates. Its tracker vendor publishes an MCP server. The studio manager adds the URL as a custom connector in Claude and sees six tools: list projects, get project, list tasks, create task, update task status and delete project.

Before anyone else connects, she reviews it. The job is reading project status, so the three read tools are allowed, create and update ask every time, and delete is off. Sign-in is each person's own tracker account, so a contractor sees only their projects. The vendor's data page says requests are logged for 30 days. The decision, "connect with conditions", goes in the team's AI rules with a review date three months out. That review took twenty minutes, and the studio found out what the AI could do before it did it.

How to connect a server safely

1. Write the job in one sentence, such as "read project status for client updates".
2. Check the publisher: the official server of the product, or a third party you can name and contact.
3. List every tool and mark it read, write or run.
4. Allow what the job needs, set "ask every time" for anything that sends, deletes, pays or publishes, and turn off the rest.
5. Record the decision with an owner and a date to review it, and look again when the tool list changes.

MCP and skills

MCP moves tools and data. Agent Skills move know-how: the steps your team follows for a task. They work together. A skill can tell the AI to use a connector's tool at step three, and an MCP server can serve a team's skills to every AI that connects. AI Skill Pod does the second: one MCP link gives Claude, ChatGPT (Pro, Business, Enterprise), Claude Code, Codex, Cursor and OpenCode the same set of team skills.

Notion's "shared brain", announced as one package on 4 November 2026, uses MCP this way too: through Notion's MCP server, ChatGPT, Claude and other agents share one person's Notion memory and the team's Notion skills, with 69 tools in Claude's directory. What Notion's shared brain does and doesn't cover walks through it, and the review steps above apply before anyone connects it.

Questions

What does MCP stand for?

Model Context Protocol. Anthropic introduced it on 25 November 2024 as an open standard for secure, two-way connections between AI tools and data sources, and donated it to the Agentic AI Foundation under the Linux Foundation on 9 December 2025.

Is MCP only for Claude?

No. ChatGPT connects to custom MCP servers (full access on Business and Enterprise/Edu, read and fetch on Pro, per OpenAI's help page), and Cursor, Codex, Claude Code and many other tools are MCP clients. OpenAI co-founded the foundation that now hosts MCP.

Is it safe to connect an MCP server?

It is as safe as the server and the permissions you give it. The specification says tools represent arbitrary code execution and that a host must get the user's consent before invoking a tool. Check who runs the server, what each tool can change, and turn off what you don't need.