MCP in one picture: plumbing, not safety

MCP, the Model Context Protocol, comes up in every agent conversation. The one-minute version:

It's an open protocol that lets an AI app talk to tools and data through one common interface. A server exposes tools (actions), resources (data) and prompts, and any client that speaks MCP can use them. Write a Slack or database connector once and it works across agent apps. MCP standardises how tools are described, called and answered. That's the plumbing.

It doesn't solve three things. It doesn't make a tool safe, because a server can still do harm or return text that tries to steer your agent. It doesn't make the model pick the right tool, so clear names and descriptions still matter most. And it doesn't make tools free: every tool description costs tokens, and a long list slows the model down and confuses it.

So install only servers you trust, give each the least access it needs, and keep the list short. Some new projects exist just to shrink tool lists. One of them, mcptoon (333 stars), claims to cut tool discovery from 71,929 tokens to 581 across 255 tools. That's its own number, so measure it yourself.

How many MCP servers is your agent connected to right now, and do you use all of them?