What Is the Model Context Protocol (MCP)? How It Connects AI to Tools

What is the Model Context Protocol (MCP)?
The Model Context Protocol (MCP) is an open protocol, a shared set of rules, that connects AI applications to external tools and data.
It lets a compatible application use tools offered by an MCP server, a program that makes those tools available. You may not need MCP if your application already calls the tool’s API, the interface it provides to other programs, directly.
An AI application, called the host in MCP, is the program you use to talk to a model, such as an assistant in your editor. Model Context Protocol sits between that application and outside tools, which are functions that can retrieve information or take an action.
Before MCP, Anthropic wrote in its announcement, every new data source needed its own custom implementation. MCP gives those connections a common interface: a server built for the protocol can be reused by compatible applications. It is one way to give an assistant access to information it cannot see on its own.
Follow one request to an issue tracker
Suppose your editor’s assistant has no access to your repository, the place where your project’s code is stored. You ask: “Which issues are open in my repository?” Issues are the tasks or reports recorded in an issue tracker, a service that manages them. Without a connection, the assistant cannot look up the answer.
A tracker vendor or developer can publish an MCP server for the tracker. You add it to your application, for example in its settings, and the application starts it or reaches it over a network. The tracker’s API is its own interface; the MCP server wraps it so compatible MCP applications can use it without their own tracker integration code.
Inside the application, an MCP client is the part that communicates with that server. The server is the separate program that offers the tracker’s functions to the application.
The application shows the model the server’s list of tools. The model picks the tool for finding issues. The specification says a person should be able to deny the call, so a host that follows it asks for your consent before the tool runs.
The client sends the request to the server; the server calls the tracker and returns the issues. The application passes the result to the model, which answers your question.
The same server can be connected to another application that speaks MCP, without writing a new server for that application. Both applications still need the connection and the necessary access to the tracker.
What a server can offer
The issue lookup is a tool, but an MCP server can offer other kinds of information too. The distinction tells you who chooses what to use:
- Tools are functions that retrieve information or take actions. The model can select a tool. The specification says there should always be a person who can deny a tool call.
- Resources are data the application can attach and manage as context, the text the model can see for a task, such as a document. The application controls their use.
- Prompts are reusable message templates, such as a request to summarize issues. The user controls their selection.
For the details of MCP architecture, roles and server offerings, including how applications connect and authorize access, read the architecture explainer.
Who created and supports MCP
David Soria Parra and Justin Spahr-Summers created MCP at Anthropic, the AI company that announced it in November 2024 to give AI systems a common way to connect to data sources. The project’s maintainers now direct its technical work; the Agentic AI Foundation, a directed fund under the Linux Foundation, is its home.
The MCP project documentation names Claude and ChatGPT, both AI assistants, and the editors Visual Studio Code and Cursor as applications that support MCP.
Microsoft’s Visual Studio Code 1.102 release notes, dated 9 July 2025, say MCP support became generally available in the editor. Support in an application does not by itself give it access to your tracker: you still need to connect and authorize a server.
If you only wanted to know what MCP is, you can stop here.
Choose MCP, a direct API call, or a terminal command
Your application can call the issue tracker’s API directly if that is the only integration it needs. You write the connection for that application, while an MCP server can offer the tracker to multiple compatible applications.
Function calling lets a model return a structured request for a function that the application or the model vendor runs. MCP tools are functions a server offers through the MCP interface. They can be used together: OpenAI’s API documentation describes remote MCP servers as an addition to the tools you make available with function calling.
A CLI, or command-line interface, lets an assistant run a command in a terminal, a text interface for operating programs. If the agent already knows a suitable tracker command and has access, that can be the simpler route.
Anthropic’s Claude Code best-practices page says that if a system has a CLI, that is usually the easiest way in, and that for systems without a CLI there is MCP.
MCP also supports remote servers with authorization, meaning a controlled grant of access, and applications can show tool calls for your approval. A direct API call or a CLI command relies on the access controls of that API or tool instead.
In its engineering post ‘Code execution with MCP’, Anthropic reported that most MCP clients at the time loaded all tool definitions into the model’s context upfront. Some clients now defer them: the Claude Code documentation says its tool search, on by default, loads only tool names and server instructions at session start.
Defining tools for function calling also uses input tokens for names, descriptions and schemas. Loading many definitions upfront can leave less room for your request and its results.
What the current version changes
The current MCP specification, version 2026-07-28, makes requests stateless: each carries the protocol version and relevant client features so the server can process it independently. Earlier versions opened a setup exchange first and kept a session; now each request carries the details the server needs.
A local server is a program on your machine; a remote server is a service reached over a network. These are the standard ways to connect; the choice affects where the server runs and how you grant it access. For implementation steps, see the guide to building an MCP server with the official SDKs.
Check a server before you connect it
For the issue tracker, check who publishes the server, what its tool descriptions say it can do, and which repositories and actions its access permits. Grant only the access the task requires. A server with permission to change issues can do more than list them; review the requested action before approving it.
The specification says hosts must obtain explicit user consent before invoking any tool, and that implementors should build consent flows. MCP cannot enforce this at the protocol level. Check that your application actually shows you the action and lets you decline. A server’s tool descriptions are also untrusted unless the server itself is trusted.
When it breaks: The arXiv preprint 2603.22489 describes risks including misleading instructions in a tool description and changes to a server after you trust it. A standard interface does not verify the server’s intentions. If you cannot establish what it can access or your application does not give you meaningful control over calls, do not connect it to the tracker.
This connection is one part of agentic and autonomous coding, where an assistant can take steps using tools as well as answer questions.
Cheat sheet
The article on one page: open it full size, save it or share it.

AI-assisted content, human-reviewed. Images AI-generated. Editorial Standards · Our Editors