MCPs: APIs But Make It Weird
What a Model Context Protocol server is, why tools show up in your agent, and when connecting one is just context clutter.
MCP stands for Model Context Protocol. In normal-person terms: it is a standard way to expose tools and resources to an LLM client.
It feels like an API, but the audience is your AI client. Instead of your code calling Notion, GitHub, a database, or a local script directly, the agent sees a list of available tools and decides when to call them.
Local vs remote
stdio MCP servers run on your machine. The client starts a process and talks to it through standard input and output. This is common for local dev tools.
HTTP/SSE style servers run remotely or as long-lived services. This is more natural for shared services and cloud-hosted tools.
What connecting means
Connecting an MCP usually means adding config that says: here is the server command or URL, here are environment variables, and here is the scope where this server should be available.
Scope matters. A global Notion MCP might be useful in your knowledge-work folder and useless noise in a codebase that has nothing to do with Notion.
The criticism
Some MCPs expose a lot of tools. If your agent is carrying 80 irrelevant tools, it can waste tokens, distract planning, and increase the chance of odd tool choices.
Good MCP hygiene is boring: install only what you need, prefer project scope over global scope, and remove servers that are not relevant to the current work.