MCP Servers and Tool Boundaries
Understand how the Model Context Protocol connects AI apps to tools and data, and where you must draw trust boundaries.
Objective
By the end of this lesson, you can explain the parts of an MCP setup, review an MCP server before connecting it, and scope its tools to the least access the task needs.
What MCP is
The Model Context Protocol (MCP) is an open protocol, introduced by Anthropic in late 2024 and since supported by many AI apps and tools. It standardizes how an AI application connects to outside tools and data.
Before MCP, every app wrote its own connector for every tool. With MCP, a tool provider writes one server, and any app that speaks MCP can use it.
The parts:
- Host: the AI application a person uses, such as a desktop assistant or a coding tool.
- Client: the piece inside the host that talks to one server.
- Server: a program that exposes capabilities. These include tools (actions the model can request), resources (data the app can read), and prompts (reusable templates).
- Transport: how they connect. Usually a local process on the same computer, or HTTP for remote servers.
What MCP does not do
MCP is plumbing. It does not decide what is safe. A server that exposes delete_all_files will offer it to the model just as readily as read_faq. Safety comes from your choices:
- Which servers you connect. Each one is code that runs with some access to your systems.
- Which tools you enable. Turn off tools the task does not need.
- What credentials the server holds. Give it a narrow account, not an admin key.
- Which actions need approval. Many hosts can ask a person before running a tool. Keep that on for anything that writes, sends, or deletes.
Risks specific to connected tools
- Untrusted servers. A server from an unknown source can run harmful code or return poisoned results.
- Tool descriptions as an attack path. The model reads each tool's description. A malicious server can hide instructions there.
- Too many tools. Dozens of connected tools confuse tool choice and widen what an injected instruction can reach.
- Mixed trust in one session. A tool that reads public web pages and a tool that sends email, in the same session, let outside text trigger outbound actions (Module 8).
Before you connect an MCP server
0 of 8 doneA national example
Quiz: MCP check
Go deeper: MCP Dashboards vs. The Old Way: A Practical Outline for Faster, Safer Analytics
Discussion
Ask. Answer. Build together.
Loading…