What Is MCP (Model Context Protocol)? Explained
What is MCP? An open standard that connects AI apps like Claude to data, tools and prompts. Here is how it works and how to start.
MCP is an open-source standard that connects AI applications to external data, tools and workflows. That is the short answer to what is MCP: one shared plug between an assistant such as Claude and the systems it needs. A server exposes actions, readable data and prompt templates, and any supporting client can use them. You can try it today by installing a server such as GitHub in Claude Code, Cursor or VS Code.
What is MCP in one sentence?
The official Model Context Protocol documentation defines MCP as an open-source standard for connecting AI applications to external systems. With it, an application such as Claude or ChatGPT can reach data sources like local files and databases, tools like search engines and calculators, and workflows like specialized prompts. That access lets the assistant fetch key information and perform tasks instead of only answering from what it already knows.
The documentation also offers an analogy: think of MCP as a USB-C port for AI applications. Just as USB-C gives electronic devices one standardized way to connect, MCP gives AI applications one standardized way to connect to outside systems. For someone who has never met the term, that is the whole idea: a common connector, so an assistant and a service do not each need a one-off integration.
The same documentation lists examples of what this enables. Agents can access your Google Calendar and Notion and act as a more personalized assistant. Claude Code can generate an entire web app using a Figma design. Enterprise chatbots can connect to multiple databases across an organization, so users can analyze data through a conversation rather than a query editor.
Where does MCP come from, and who maintains it?
MCP is published as an open standard at modelcontextprotocol.io, which hosts the documentation, the specification, extensions, a registry, proposals called SEPs, and a community section. As of October 2026, the site labels version 2026-07-28 as the latest. Because the specification is versioned, it is worth checking which version a client or server targets before you assume two tools will work together.
The reference servers are maintained alongside the protocol. The official GitHub MCP server is described as run by the Anthropic and Model Context Protocol team, and it is distributed as the npm package @modelcontextprotocol/server-github. Community forks also exist for people who want broader coverage, such as Actions, projects and packages support, but they are separate projects with their own maintainers.
It helps to keep the layers apart. MCP is the protocol, the shared language. A server is a program that speaks it for one system, such as GitHub. A client is the application, such as Claude Code or Cursor, that connects to servers on your behalf. A directory like this one catalogues servers; it is not the protocol itself.
What are tools, resources and prompts in MCP?
A server can expose three kinds of capability. Tools are actions the agent can call, such as opening an issue. Resources are data the agent can read. Prompts are reusable templates that package a workflow. The protocol documentation describes the same split in plain terms: data sources, tools and workflows such as specialized prompts. Our deeper write-up is MCP Tools, Resources, and Prompts Explained.
Tools are the easiest to picture. The GitHub server's page lists nine tools, including get_repository, list_issues, create_issue, list_pull_requests, create_pull_request, get_file_contents and create_or_update_file. Each one has named inputs, such as an owner and a repo, and some are marked required or optional. An agent reads those definitions and decides which tool fits the task.
The distinction matters when you choose or build a server. If the job is to do something, such as create a pull request, it is a tool. If the job is to let the agent read something, such as a file or a record, it is a resource. If you want a repeatable instruction like a review checklist, it is a prompt. Many servers lean mostly on tools.
What does MCP change for a developer?
The practical change is that the agent calls the service directly. Notes on the GitHub server describe PR review automation: the agent calls list_pull_requests, fetches the diff and leaves contextual comments. They describe issue triage the same way, for example looking at the oldest open issues, grouping them by area and suggesting labels. The alternative is copy-pasting issue text into a chat window by hand.
It also moves some responsibility onto you. Those notes warn that fine-grained GitHub tokens have surprising scope limits: the all-repositories toggle does not include private repos by default, so an empty list_pull_requests result may simply mean the token lacks that repo. They recommend the smallest scope, read-only repo and issues access, and upgrading to write access only with a human approving each step.
Cost and safety are part of the picture too. The same notes say create_or_update_file commits through the API, with no GPG signing, no pre-commit hooks and possibly no CI run, so anything beyond small docs fixes should go to a branch and a pull request, never straight to main. They also say reading a 1k-line file with get_file_contents costs roughly 4k tokens, while issue and PR reads stay lean.
Which clients support MCP today?
The protocol documentation names Claude and ChatGPT as examples of AI applications that can connect through MCP. For coding work, the install instructions for the GitHub and VS Code servers target Claude Code, Claude Desktop, Cursor, Windsurf, and VS Code with GitHub Copilot, and they also say the servers work with any MCP client. Support varies by client and version, so confirm it in your client's own documentation before you rely on it.
VS Code deserves a note because the name appears in two roles. As a client, it can host MCP servers; our guide How to Add an MCP Server to VS Code walks through that setup. As a server, the VS Code MCP server reads workspace structure and diagnostics, so another agent can see the project you have open. Its profile documents no authentication setup, so check the upstream documentation before installing.
Where a server fits is decided by what it exposes, not by which client you use. The GitHub server suits teams whose code and pull request comments live in GitHub. Its listed alternatives are GitLab MCP for GitLab teams, Gitea MCP for self-hosted setups, and Linear MCP for sprint and cycle planning rather than code.
How do I get started with MCP?
Pick one client you already use and one server that matches a real task. A sensible first pair is a coding client and the GitHub server, because the read tools are low risk. The steps look like this:
- Choose a client, such as Claude Code, Cursor or VS Code.
- Install the server with
mcpizy install github, which writes the config to your client, or run it directly withnpx -y @modelcontextprotocol/server-github. - Create a fine-grained token with read-only repo and issue access, and select each private repo explicitly.
- Ask the agent for a read-only job, such as summarizing the oldest open issues.
- Add write access only when a human approves each change.
For a full walkthrough with Claude Code, read How to Use GitHub MCP with Claude Code. If you prefer VS Code, the install command is mcpizy install vscode, or npx -y vscode-mcp-server to run it directly. Whichever you choose, start with one server, confirm it behaves as expected, and add more only once the first one earns its place.
MCP FAQ: quick answers
Is MCP free? The protocol documentation describes MCP as an open-source standard. The GitHub server's page lists it as free with nine tools, and the VS Code server's page lists it as free and open-source. The official pages we drew on do not state a price for any other server, so check each vendor, since a server may call a paid service behind it.
Do I need to write code to use MCP? Not to use an existing server. A single install command, such as mcpizy install github, writes the configuration into your client. Writing your own server is a separate task that the protocol documentation covers, and it is not needed for a first try.
Is it safe to give an agent access? It is as safe as the permissions you grant. Use the smallest token scope, begin read-only, route changes through branches and pull requests, and keep a human approving writes. Remember that the protocol is a standard for connecting; what a server can do depends on that server and on the credentials you hand it.