Blog ·
MCP, the Model Context Protocol, is an open standard for connecting AI apps to outside tools and data. An app that speaks MCP, like Claude, ChatGPT, Cursor or VS Code, can plug into any MCP server: one for GitHub, one for your database, one for a remote computer. The AI can then use what that server offers, such as reading a file, querying a table or running a command.
MCP's own docs compare it to a USB-C port for AI: one standard plug instead of a different cable for every device.
Where MCP came from
Anthropic introduced MCP as an open standard in November 2024. Within a year it was supported by most major AI apps and coding tools, including ones from OpenAI, Google and Microsoft. In December 2025 Anthropic donated it to the Agentic AI Foundation, a new fund under the Linux Foundation co-founded by Anthropic, Block and OpenAI, so no single company controls it.
The problem MCP solves
Before MCP, every AI app needed its own integration for every service. Ten apps and a hundred services meant up to a thousand custom connectors, each built and maintained separately.
With MCP, a service builds one MCP server and it works in every app that supports MCP. An app adds MCP support once and gets every server. Ten apps plus a hundred services means a hundred and ten pieces, not a thousand.
How MCP works
There are three roles:
- The host: the AI app you use, such as Claude Desktop, Claude Code, Cursor or ChatGPT.
- The client: the part of the host that holds a connection to one server.
- The server: a program that offers capabilities to the AI. It can run on your machine or on the internet.
A server can offer three kinds of things:
| Offers | What it is | Example |
|---|---|---|
| Tools | Actions the AI can take | run_command, create_issue, send_message |
| Resources | Data the AI can read | A file, a database row, a document |
| Prompts | Ready-made instructions a user can pick | “Summarise this pull request” |
When you ask your AI app something, this happens behind the scenes:
- The app connects to its MCP servers and asks each one what tools it has. Each tool comes with a name, a description and the inputs it takes.
- The app passes those descriptions to the model along with your message.
- The model decides a tool would help and asks to call it, with specific inputs.
- The app sends the call to the server, often after asking you to approve it.
- The server does the work and returns the result, which goes back to the model.
- The model uses the result to answer you, or calls another tool.
Servers on your own machine usually talk to the app over standard input and output. Remote servers use HTTP, often with an API key or an OAuth sign-in.
A worked example
Say you've connected Claude Code to the Cubicles MCP server, which controls cloud computers. Adding it is one command:
claude mcp add --transport http cubicles https://api.cubicles.cloud/api/v1/mcp \
--header "Authorization: Bearer cub_…"Now you ask: “Is my research computer running? If it is, show me its screen.” Claude Code sees the server offers list_desktops and take_screenshot. It calls the first, finds the computer, calls the second, and shows you the screenshot. You never wrote any code, and Claude Code never needed a Cubicles-specific feature.
What people connect over MCP
- Code and issues: GitHub, GitLab, Linear and Jira, so a coding agent can read tickets and open pull requests.
- Data: Postgres, SQLite and data warehouses, so the AI can query instead of guessing.
- Documents and chat: Notion, Google Drive and Slack, for answers grounded in your team's own notes.
- Browsers: servers that drive a real browser, so the AI can test a web app or read a page that needs JavaScript.
- Computers: servers like Cubicles' that run commands, read and write files and take screenshots on a remote machine. See give Claude Code, Cursor or Codex a cloud computer with MCP.
- Monitoring: error trackers and logs, so the AI can read the stack trace itself.
MCP vs an API
MCP doesn't replace APIs. Most MCP servers are a thin layer over an API that already exists. What MCP adds is everything a model needs to use that API on its own:
| A regular API | An MCP server | |
|---|---|---|
| Written for | Developers writing code | AI models choosing tools at runtime |
| Discovery | Read the docs | The app asks the server what it offers |
| Descriptions | For people | For the model, so it knows when to use each tool |
| Integration work | Per app, per API | Once per server, then it works in every MCP app |
If you're writing a script, call the API. If you want an AI app to use the service by itself, give it an MCP server.
MCP vs an agent
MCP isn't an agent. It's how an agent, or any AI app, reaches tools. Claude Code with an MCP server is an agent using a tool. OpenClaw running on its own computer is an agent with tools built in. Both can be useful together: an agent on a remote computer does long jobs, and MCP lets the assistant in your editor check on it.
The security risks
An MCP server is software that can act on your behalf, so treat it like any other software you install. The main risks:
- Untrusted servers. A local MCP server runs code on your machine. A malicious one can do anything you can.
- Poisoned tool descriptions. A server's tool descriptions go straight to the model, and can hide instructions in them.
- Prompt injection through results. A web page, issue or email that a tool returns can contain instructions aimed at the AI.
- Over-broad access. A server connected with your admin token can do everything your admin account can.
The rules that keep you safe:
- Install servers only from sources you trust, ideally the service's own official server.
- Use scoped, expiring tokens, never your main account's full-access key.
- Keep approvals on for actions that change things, like writes, sends and deletes.
- Point powerful servers at a separate machine. A server that runs shell commands is far safer aimed at a disposable cloud computer than at your laptop.
The AI agent security checklist applies to MCP setups too.
Try it
The quickest way to feel what MCP does is to give your AI app a computer of its own. With a free Cubicles computer and an API key, Claude Code, Cursor, VS Code, Codex or Claude Desktop can run commands, read and write files and take screenshots on it. The MCP docs list every tool and the setup for each app.
Frequently asked questions
What does MCP stand for?
Model Context Protocol. It's an open standard for connecting AI apps to outside tools and data.
Who created MCP?
Anthropic introduced it in November 2024. In December 2025 it was donated to the Agentic AI Foundation under the Linux Foundation, co-founded by Anthropic, Block and OpenAI.
Is MCP free to use?
Yes. The protocol is open, and its specification and official SDKs are open source. Individual MCP servers may connect to paid services.
Does ChatGPT support MCP?
Yes. ChatGPT, Claude, Cursor, VS Code, Codex and many other apps can connect to MCP servers, though which plans and features support it varies by app.
Is MCP the same as an API?
No. Most MCP servers sit on top of an API. MCP adds a standard way for AI apps to discover tools and understand when to use them, so one server works in every MCP app.
Are MCP servers safe?
They're as safe as the software and access you give them. Install servers from trusted sources, use scoped tokens, keep approvals on for actions that change things, and point servers that run commands at a separate machine.