Claude Code MCP setup and usage rules

Claude Code MCP setup and usage rules

I have been experimenting with various MCP servers for Claude Code. After over a year of trial and error, I have narrowed my setup to four active servers across two machines: two on my work laptop and two on my Mac mini. I also removed several others that didn't fit my workflow.

On my work laptop, Context7 is a staple. This MCP feeds the latest library and framework documentation to the model, preventing Claude from writing outdated API code based on its training cutoff. To optimize this, I added specific rules to my global CLAUDE.md instructing Claude to use Context7 automatically for code generation, configuration, migration, or example verification. Local documentation in the repository takes priority, with Context7 used for supplementation. Crucially, if a search fails or documentation is missing, Claude must explicitly state this before proceeding; this prevents the model from silently reverting to its own training data. I also added a rule that if a code analysis agent makes a claim about how a library works, Claude must verify it via Context7 or the web before applying changes.

The second tool on my laptop is CodeGraph, which uses tree-sitter to parse code into a graph of symbols, call relationships, and files. The key is separating structural queries from literal searches. I configured it so that questions about definition locations, callers, change impact, or signatures use the graph tool, while literal strings, comments, or log messages are handled by grep. To avoid inefficiency, I instructed Claude to answer "How does X work?" using one context tool call and two or three exploration calls, rather than looping through sub-agents and grep. I also established boundaries: since the graph updates approximately 0.5 seconds after a file is saved, Claude should not query it again in the same turn immediately after a modification. Furthermore, because graph results are parsing-based, Claude should not use grep to re-verify them. Without these rules, Claude would often run the graph tool and then immediately run grep to be sure, doubling the processing time.

I eventually removed my browser automation MCP because the connection died whenever the session changed, forcing constant restarts. It even left a note in its memory stating, "session dies between conversations, restart required." Ultimately, it was more stable to control the browser using direct scripts.

The setup on my Mac mini is different. One MCP handles Discord; a plugin allows me to use Claude Code as a Discord bot with tools for replying, reacting, and downloading attachments. In fact, I am writing this post based on instructions received through that interface. The other is a Chrome extension. While using it today to check Search Console, I found that screen captures kept timing out because the page was constantly loading; only reading the body text via JavaScript worked. Whenever I find a reliable method, I save it to the memory.

Reviewing my CLAUDE.md and work logs, it is clear that the real effort isn't in attaching the MCPs, but in defining the rules. Without those few lines in CLAUDE.md explaining when to use a tool, when to stop, and how to report failure, the tools are either ignored or used redundantly.

MCP is less about the connection and more about the rules.

Related posts

This post is the English edition of a Korean write-up: 원문 보기

Comments

Post a Comment

Popular posts from this blog

npm command not found on Windows: fix the PATH

Image
I recently hit a wall while configuring my development environment on Windows. It's one of those subtle issues that can waste an afternoon if you don't know what to look for. Here's exactly how I diagnosed and fixed a 'command not found' error after a successful global install. The Problem I decided to add a markdown processing tool to my toolkit by running: npm install -g markdown-it-cli The terminal output confirmed the installation succeeded, displaying a message about added packages. Satisfied, I tried to use it by typing markdown-it-cli in the command line. Windows immediately responded with: 'markdown-it-cli is not recognized as an internal or external command, operable program or batch file.' This error means Windows looked through every folder in its search list and couldn't find an executable with that name. Attempts and Failures My initial reaction was to suspect the package itself might be broken. I tried running it using npx , wh...
Read more

Tailscale without sudo: what userspace mode actually costs you

At first, I installed Tailscale on macOS knowing I did not have admin rights to write to system directories. The CLI package installed cleanly enough, landing entirely within my user space without triggering any permission errors or requiring elevated prompts. The installation process felt seamless, but the friction appeared immediately during usage. Every single command I typed required appending --socket=$HOME/.tailscale/tailscaled.sock to talk to the daemon. I found myself adding this flag to status checks, key generation, and connection queries across multiple sessions. The repetition was exhausting for a tool I expected to run in the background, turning routine operations into a chore. I eventually wrote a shell function to wrap this flag, allowing me to call tailscale directly without typing the socket path every time. This workaround saved my sanity but highlighted that userspace mode demands extra configuration just to use the CLI, and any automation scripts I maintained had...
Read more

Claude CLI 401 Unauthorized Refresh Token Issue

Symptom: claude commands fail with 401 Unauthorized and the message "refresh token was already used. Please log out and sign in again." Cause: The refresh token expired after being consumed by another session or a background process. Refresh tokens are single-use credentials. Fix: Run claude auth login to refresh authentication without running a logout command first. Why claude login alone did not clear the 401 I was running a pipeline command when every subsequent CLI call failed. The terminal returned this exact output: Error: 401 Unauthorized Message: "Authentication failed: refresh token was already used. Please log out and sign in again." Code: "AUTH_ERROR" The error text explicitly instructs me to log out. I spent roughly ten minutes searching the terminal for a logout function. The prompt only showed AUTH_ERROR . I assumed clearing old sessions was required first. That assumption blocked progress. The instructions in the error message ...
Read more