MCP
MCP stands for Model Context Protocol. It is an open way for agents to connect to external capabilities such as documentation sources, databases, internal tools, issue trackers, or custom systems.
Quick Take
Section titled “Quick Take”Use MCP when the agent needs structured, repeatable access to a system. Do not use it just because it sounds advanced.
Why It Matters
Section titled “Why It Matters”Without MCP, you often paste context into prompts or rely on generic web browsing. With MCP, the agent can get structured access to the exact system it needs.
That usually means:
- less prompt stuffing
- more reliable context retrieval
- better tool discoverability
- workflows that are easier to share across teams
When To Use It
Section titled “When To Use It”MCP is a good fit when the agent needs repeatable access to:
- official documentation
- internal knowledge systems
- tickets and project data
- controlled tools that should not be re-described in every prompt
The Basic Pieces
Section titled “The Basic Pieces”| Piece | What it does |
|---|---|
| Tools | functions the agent can call |
| Resources | data the agent can read as context |
| Prompts | reusable prompt templates from the server |
| Servers | the processes exposing those capabilities |
Common Coding MCP Servers
Section titled “Common Coding MCP Servers”| Server type | Why people add it |
|---|---|
| GitHub | issues, PRs, code search, repo actions |
| Browser automation | screenshots and UI verification |
| Filesystem | carefully scoped file access |
| Database | read-only inspection and debugging |
| Web search | current docs and release changes |
Safety Basics
Section titled “Safety Basics”- Only connect servers you trust.
- Assume external content can contain prompt injection.
- Give the agent access to the smallest useful surface area.
- Prefer read-only integrations unless write access is clearly justified.
The Practical Rule
Section titled “The Practical Rule”If you keep pasting the same docs or instructions into prompts, that is a sign the information probably wants to live in a context file, a skill, or an MCP server instead.
Starter .mcp.json
Section titled “Starter .mcp.json”{ "mcpServers": { "github": { "command": "npx", "args": ["-y", "@modelcontextprotocol/server-github"], "env": { "GITHUB_TOKEN": "${GITHUB_TOKEN}" } }, "playwright": { "command": "npx", "args": ["-y", "@playwright/mcp-server"] } }}Cost And Noise Rule
Section titled “Cost And Noise Rule”Every server adds more capability, but also more surface area and usually more context overhead. Keep only the servers that help with the task at hand.