Foundations 4 min
MCP — how AI plugs into your tools
The standard that lets agents talk to your filesystem, Slack, GitHub, Postgres, etc.
MCP (Model Context Protocol) is a standard for letting an AI assistant reach into the rest of your stack — your filesystem, your Slack, your Linear board, your database, your browser. Without MCP, the AI only knows what's in the current chat. With MCP, it can read a real-time spreadsheet, post to Slack, search Postgres, or click around a webpage.
The 30-second mental model
- An MCP server exposes a set of tools (e.g. "read file", "post message", "run SQL").
- An MCP client is your AI tool (Claude Code, Cursor, Continue, etc.) that knows how to call those tools.
- You install the server, configure the client to find it, restart, and now your AI can use it.
What people actually use it for
| MCP server | What it unlocks |
|---|---|
| Filesystem | "Summarise all PDFs in ~/Downloads" |
| GitHub | "Open a PR for this branch with a good description" |
| Slack | "Post the deploy result to #ops" |
| Postgres / SQLite | "How many active users last week?" |
| Linear / Jira / Asana | "What's blocked in my queue?" |
| Playwright / browser | "Click around the staging site and report any 500s" |
| Google Drive / Notion | "Update the spec doc with these decisions" |
Setup pattern (the same for almost any MCP)
- Find the MCP server. Most are at github.com/modelcontextprotocol/servers or as npm/pypi packages.
- Add it to your AI client's MCP config. In Claude Code:
claude mcp add <name> <command>. In Cursor: Settings → MCP → Add. - Restart the client.
- Ask the AI: "What MCP tools do you have?" — it'll list them.
Why it's a big deal. Before MCP, every AI tool had its own incompatible plugin
system. With MCP, write a server once, use it from any client. The list of available servers
grows by the week.
Where MCP shows up in this tutorial
- Coding CLIs (Claude Code, Cursor, Continue) — primary clients.
- Open-source agents (Hermes, OpenClaw, Goose) — most support MCP.
- Chat apps — partial support, growing. Claude.ai supports remote MCP in some plans.
Who made MCP?
Anthropic open-sourced the spec in late 2024. OpenAI, Google, Microsoft, and the major editor
vendors adopted it through 2025/26. It's a community standard now, governed openly.
Are MCP servers safe?
A server runs on your machine (or wherever you put it) and has whatever access you give it.
Treat third-party MCP servers like any code: prefer the official ones from
modelcontextprotocol/servers,
scope credentials tightly, read the code for anything that touches sensitive data.
Can I write my own MCP server?
Yes — there are SDKs in TypeScript, Python, Go, Rust, Kotlin, C#. A trivial server is ~40 lines.
See modelcontextprotocol.io.