DEV Community

hyad
hyad

Posted on

MCP Servers Changed How I Use Claude Code — Here's My Setup

#ai #programming #devtools #productivity

Claude Code out of the box reads files and runs terminal commands. That's useful, but it's a fraction of what it can do.

MCP (Model Context Protocol) servers extend Claude Code with direct access to databases, GitHub, browsers, Google Sheets, and dozens of other services. Once I set them up, my workflow changed completely — no more copy-pasting data between tools.

Here's how I configured my MCP stack and which servers are actually worth using.

What MCP servers do

Think of them as plugins. Each server exposes tools that Claude Code can call during any task. You don't need to tell it "use the GitHub MCP server" — you just say "create a PR for these changes" and it figures out which tool to call.

The architecture:

  1. You configure a server (locally or remotely)
  2. Claude Code discovers its tools at startup
  3. During tasks, it calls tools as needed
  4. Results flow back into context

It feels native. That's the key insight.

Setting up your first server

One command:

claude mcp add github --scope project -- npx @modelcontextprotocol/server-github
Enter fullscreen mode Exit fullscreen mode

The --scope flag matters:

  • user — available in all projects
  • project — only in the current project

Or edit .mcp.json directly:

{
  "mcpServers": {
    "github": {
      "command": "npx",
      "args": ["@modelcontextprotocol/server-github"],
      "env": {
        "GITHUB_PERSONAL_ACCESS_TOKEN": "ghp_..."
      }
    }
  }
}
Enter fullscreen mode Exit fullscreen mode

The servers I actually use daily

GitHub

Manage issues, PRs, code search without leaving Claude Code. "Find all open PRs that touch the auth module" just works.

Playwright (browser automation)

Claude Code controls a browser — navigates pages, clicks elements, fills forms, takes screenshots. I use this for automated testing and visual verification after deploys.

Database (Turso/SQLite)

Query your database directly. Instead of writing SQL in a separate client and pasting results, Claude Code runs queries and acts on the data inline.

Memory server

A structured knowledge graph that persists across sessions. Facts survive between conversations. This pairs well with CLAUDE.md configuration for project-level context.

My production .mcp.json 

{
  "mcpServers": {
    "github": {
      "command": "npx",
      "args": ["@modelcontextprotocol/server-github"],
      "env": { "GITHUB_PERSONAL_ACCESS_TOKEN": "ghp_..." }
    },
    "playwright": {
      "command": "npx",
      "args": ["@anthropic-ai/mcp-server-playwright"]
    },
    "memory": {
      "command": "npx",
      "args": ["@anthropic-ai/mcp-server-memory"]
    }
  }
}
Enter fullscreen mode Exit fullscreen mode

Start with 2-3 servers. Each one adds tool definitions to Claude Code's context window, so more servers = more context consumption. Add what you need, disable what you don't.

The context window trap

This is the thing nobody warns you about. Every enabled MCP server adds tool definitions to Claude Code's system prompt. Ten servers with dozens of tools each means significant context overhead on every request.

This directly impacts rate limits. More context per request = fewer requests before you hit limits.

To manage it:

  • Disable unused servers: claude mcp disable server-name
  • Use project-scope for project-specific servers
  • Use user-scope only for universally useful ones

Before and after MCP

Without MCP With MCP
Query DB separately → copy → paste into chat "Find users who signed up this week" — auto-queries
Open GitHub → create PR → copy URL → tell AI "Create a PR for these changes" — handles everything
Write script → run → read output → paste back Tool calls happen inline, automatically

The productivity difference is real. MCP eliminates the copy-paste-context-switch loop.

The full picture

MCP servers are one layer of the Claude Code configuration stack. The full stack is:

  1. CLAUDE.md — project context and conventions
  2. Persistent memory — cross-session continuity
  3. Custom skills — repeatable procedures
  4. Hooks — automated guardrails
  5. MCP servers — external tool access

Each layer compounds. I wrote a complete MCP servers guide covering every server worth using and how to build a production stack.

If you want the full configuration stack pre-built — skills, memory, hooks, agents, and MCP setup guides — Claudify packages it all into one install.

What MCP servers are you using? Drop your setup in the comments.

Top comments (0)