MCP servers for Claude Code

A Model Context Protocol config that earns its place, plus the traps that cost real time.

MCP turns Claude Code from a code generator into something that can read your repo, query a database, drive a browser and open a PR. MCP wiring is independent of the gateway — it changes what tools exist, not where model calls go.

A .mcp.json worth running

Copy{
  "mcpServers": {
    "filesystem": {"command": "npx", "args": ["-y", "@modelcontextprotocol/server-filesystem", "."]},
    "git":        {"command": "uvx", "args": ["mcp-server-git", "--repository", "."]},
    "github":     {"command": "npx", "args": ["-y", "@modelcontextprotocol/server-github"],
                   "env": {"GITHUB_TOKEN": "..."}},
    "fetch":      {"command": "uvx", "args": ["mcp-server-fetch"]},
    "playwright": {"command": "npx", "args": ["-y", "@playwright/mcp@latest"]}
  }
}

Drop it in the repo root; Claude Code detects it and asks you to trust it once.

The traps

• The trust prompt is a feature. Claude Code refuses to load project MCP servers until you approve them, per project. That is the security boundary, not a bug.
• Give a database server a read-only role. An MCP server executes whatever SQL the model emits; a read-only role makes a hallucinated DROP physically impossible.
• Scope tokens narrowly. For GitHub, repo plus read:org is enough. Do not hand it an all-scopes classic PAT.
• MCP output is context. A 200-row result or a full page fetch is resent on every subsequent turn. Large tool outputs are the main reason agentic sessions get expensive.