If you’ve used Claude, Cursor, or Ollama recently, you may have seen “MCP” mentioned without much explanation. MCP — the Model Context Protocol — is the open standard that lets AI models connect to your tools, files, and data sources safely and consistently. It launched in late 2024 and by 2026 has become one of the fastest-growing pieces of AI infrastructure, with over 10,000 active servers and adoption from Claude, ChatGPT, Cursor, Gemini, and Microsoft Copilot. If you’re running local AI or building anything with it, it’s worth understanding.
What is MCP?
MCP is an open protocol that standardises how AI models talk to external systems. Before MCP, every AI tool that wanted to read your files, query a database, or call an API needed its own custom integration — built by hand, one connection at a time. MCP replaces that with a single, shared way for a model to discover what tools and data are available, and to use them.
Think of it like a USB port for AI. Before USB, every device needed its own cable and driver. USB gave everything a common connector. MCP does the same for AI models and the tools they need to be useful — your codebase, your database, your home automation system, your file storage.
An MCP setup has two sides: a client (the AI application — Claude Desktop, Claude Code, Cursor, or your own app) and a server (a small program that exposes a specific capability — “read files in this folder,” “query this database,” “control these smart home devices”). The client talks to as many servers as you connect, and the model decides which tools to use based on what you ask it to do.
Why MCP Matters (Especially If You Self-Host)
For anyone already running local AI — Ollama, Open WebUI, LocalAI — MCP is the missing piece that turns a model that can only chat into one that can actually act:
- No more one-off integrations. A single MCP server (say, one that reads your Postgres database) works with any MCP-compatible client, not just one app.
- Keeps sensitive access local. Self-hosting your own MCP servers means your database credentials, file access, and internal API keys never pass through a third party’s infrastructure — the same privacy logic that makes Ollama appealing in the first place.
- Composability. Run a filesystem server, a database server, and a home automation server together, and a single AI client can reason across all three in one conversation.
- It’s becoming the default. With near-universal client support now (Claude, Cursor, VS Code, ChatGPT, Copilot) and a large official registry of ready-built servers, MCP is quickly becoming the standard way AI tools connect to anything, rather than a niche feature.
Getting Started with MCP
The fastest way to try MCP is through Claude Desktop or Claude Code, both of which support it natively:
- Open Claude Desktop’s settings and find the “Developer” or “MCP Servers” section.
- Add a server from the official MCP servers repository — the filesystem server is a good first test, since it just needs a folder path.
- Restart the client. The new tool appears automatically, and Claude will use it when relevant to your request.
For self-hosting your own MCP server rather than using someone else’s, both Python and TypeScript have official SDKs, and a minimal server is a small amount of code — often under 50 lines for a single tool. We cover the full self-hosting process, including running one alongside Ollama on a home server, in the dedicated guides linked below.
What Can You Actually Do With It?
Practical uses that are already common by mid-2026:
- Database access. Let a model query your PostgreSQL or SQLite database directly and answer questions about your own data, with no data leaving your network if the server is self-hosted.
- Codebase awareness. Cursor and Claude Code use MCP-style access to read and reason about an entire project, not just the file you have open.
- Home automation. MCP servers exist for Home Assistant, letting you control smart home devices through natural conversation instead of a dedicated app.
- Private API access. Wrap an internal company API in an MCP server so an AI assistant can use it without the API being exposed publicly.
- Local AI agents. Combine Ollama with an MCP server to build agents that run entirely on your own hardware — no cloud model, no cloud tools.
Limitations to Know
MCP is maturing fast, but it’s not without rough edges:
- Session management for remote (non-local) MCP servers is still an evolving area — self-hosting at scale requires some architectural thought, not just installing a package.
- Not every MCP server is well-maintained or secure. Installing one means giving a model access to whatever that server exposes, so vetting matters — a full guide on vetting MCP servers before installing is coming soon on this site.
- The specification itself is still evolving; a server built against an older version may not support newer features like streamable HTTP.
Next Steps
If you’re already running Ollama, Claude, or Cursor, MCP is very likely something you’ll end up using whether you deliberately set it up or not — client support is now close to universal. Spend twenty minutes connecting a single MCP server (the filesystem one is the easiest first step) and see what changes when your AI assistant can actually look at something rather than just talk about it.
