You want to tell an AI “read this file”, “pull that record from my database”, “close this issue”. The model cannot do any of it on its own, because it has no access to your machine, your database or your accounts.
MCP exists to close exactly that gap. It puts a standard connection layer between the AI application and the outside world 🔌
What Is MCP?
MCP (Model Context Protocol) is an open-source standard for connecting AI applications to external systems.
The common analogy: MCP is the USB-C port of AI. Before USB-C every device had its own cable. Before MCP every AI tool was wired into every service separately. Now there is one standard.
Anthropic created MCP and then donated it to the Agentic AI Foundation under the Linux Foundation in December 2025, backed by AWS, Google, Microsoft, Salesforce and Snowflake.
The numbers show how fast it moved: monthly downloads reached 97 million in March 2026, up from 2 million at launch, a 48x increase. More than 10,000 public servers are registered.
What Is MCP Used For?
Concrete examples. With MCP:
- An agent can reach your Google Calendar and Notion and behave like a genuinely personal assistant
- Claude Code can generate an entire web app from a Figma design
- An enterprise chatbot can connect to several company databases so anyone can analyze data by chatting
- An AI can build 3D designs in Blender and send them to a 3D printer
The common thread: the model stops just answering and starts doing the work.
One boundary worth knowing: MCP is only the connection protocol. It does not decide how the model uses that information or when to call which tool. That part is up to the application.
How Does MCP Work?
There are three pieces, and the names confuse people at first:
| Piece | What it is | Example |
|---|---|---|
| Host | The AI application you use | Claude Code, Claude Desktop, VS Code, Cursor |
| Client | The component inside the host that talks to one server | One per server |
| Server | The program exposing tools and data | Filesystem, GitHub, a database |
The one detail that matters: the host opens a separate client for each server. Connected to three servers means three clients inside the host.
The flow works like this:
- Connect. The host opens a client for every server you added and establishes the connection.
- Discover. The client asks the server “what can you do?”. The server returns its list of tools, what each one does and which parameters it expects.
- Decide. The model decides it needs one of those tools to finish the job.
- Call. The host calls the tool with the right parameters. At this point it is the server running the code, not the model.
- Result. The server does the work and returns the result. The host feeds that back to the model as context, and the model builds its answer around it.
Step 2 is where MCP gets its real power. The tool list is not fixed, it is requested at runtime, so when a server gains a new capability you do not have to update the client.
What Does an MCP Server Offer?
Servers can expose three kinds of things:
| Type | What it does | Example |
|---|---|---|
| Tools | Functions the model can call to take action | Write a file, call an API, query a database |
| Resources | Data that gives the model context | File contents, a database record, a schema |
| Prompts | Ready-made templates | System prompts, worked examples |
In practice tools get used the most. Picture a database server: it can offer a tool for querying, a resource holding the schema, and a prompt showing how to use those tools well.
Local or Remote?
There are two kinds of server, and knowing the difference helps when you set one up:
| Local server | Remote server | |
|---|---|---|
| Where it runs | On your machine | On the service's own infrastructure |
| Technical name | stdio | Streamable HTTP |
| Speed | No network hop, the fastest | Depends on your connection |
| Authentication | Not needed | Needed, usually OAuth |
| Typical use | Your files, a local database | Cloud services like GitHub or Sentry |
A filesystem server runs on your machine because it needs to reach your files. Sentry’s server runs on Sentry’s own infrastructure.
What Is Out There? Common MCP Servers
There are over 10,000 registered servers. The densest category is developer tooling, because the first adopters were developers trying to make their coding assistants more capable.
The types you will run into most:
- Filesystem: reading and writing your local files. Usually the first one people install
- GitHub: opening issues, reviewing PRs, searching repos
- Databases: querying services like Postgres and Supabase
- Error tracking: pulling error reports from tools like Sentry
- Browser automation: opening pages, taking screenshots, filling forms
- Notes and docs: reaching Notion, Google Drive and similar
- Design: pulling components and styles out of Figma files
MCP servers get access to sensitive data. Every server you install opens a door for the AI into your machine or your accounts.
In team and enterprise settings, only add servers you trust. For remote servers prefer an OAuth flow over embedding a raw API key.
How to Add MCP to Claude Code
The practical part. Claude Code is an MCP host, and adding a server is one command:
claude mcp add <name> -- <command>
For a server distributed through npm:
claude mcp add weather -- npx -y weather-mcp
To see what you have added:
claude mcp list
Nothing else is required after that. When Claude needs one of that server’s tools for the job at hand, it calls it on its own.
Should You Write Your Own Server?
Not everyone needs to. The decision is simple:
| Situation | What to do |
|---|---|
| The integration you need already exists | Install the ready-made server |
| You need access to an internal API or database | Write your own server |
| One project, one-off job | No server needed, a script will do |
| You will use the same integration across several AI tools | Write a server |
That last row is MCP’s real value: write it once, and it works in Claude, Cursor and VS Code without a code change.
Most MCP guides you will find online describe an older version of the protocol. Sampling and logging were removed.
If you are writing a new server, do not build on either: integrate directly with your LLM provider’s API, and write logs to standard error instead.
Frequently Asked Questions
What is MCP? MCP (Model Context Protocol) is an open-source standard for connecting AI applications to systems like files, databases and external services. Anthropic built it and donated it to the Agentic AI Foundation under the Linux Foundation in December 2025. The common analogy is that it is the USB-C port of AI.
How does MCP work? The AI application you use, the host, opens a client for each server and connects. The client asks the server which tools it offers, and the server returns the list along with the parameters each one expects. When the model needs a tool, the host calls it, the server runs the code and returns the result, and the host feeds that back to the model as context.
What is MCP used for? It lets an AI stop merely answering and start doing work. An assistant that reaches your calendar and notes, a coding tool that builds an app from a Figma design, or a chatbot that queries company databases are all made possible by MCP.
What does an MCP server offer? Three kinds of things: tools, which are functions the model can call to take action; resources, which are data sources that give the model context; and prompts, which are ready-made templates. Tools are what get used most in practice.
What is the difference between a local and a remote MCP server? A local server runs on your own machine, has no network hop and needs no authentication, which suits jobs like reaching your files. A remote server runs on the service’s own infrastructure, is reached over the internet and usually requires authentication, typically OAuth.
How do I add an MCP server to Claude Code? Run claude mcp add <name> -- <command> in your terminal. For a server distributed through npm that looks like claude mcp add weather -- npx -y weather-mcp. You can see what you have added with claude mcp list. Nothing else is needed afterwards, Claude calls the tool itself when it needs it.
Which MCP servers are available? There are over 10,000 registered servers. The most common ones cover the filesystem, GitHub, databases like Postgres and Supabase, error tracking tools like Sentry, browser automation, document services like Notion and Google Drive, and design tools like Figma.
Should I write my own MCP server? If the integration you need already exists, installing the ready-made server is enough. Writing your own makes sense when you need access to an internal API or database, or when you will use the same integration across more than one AI tool.
Summary
MCP in one sentence: write the integration once, and every AI application that speaks the protocol can use it.
What to remember:
- Your host application opens a separate client for each server
- Servers offer tools, data and templates, and tools get used most
- Local servers run on your machine, remote servers on the service’s own infrastructure
- Adding one to Claude Code is a single command
- Every server you install is an access door, so do not add ones you do not trust
For the autonomous systems built on top of these tools see the agentic AI post, and for using MCP servers day to day see the Claude Code guide 👀