The best MCP servers for Claude Code (and the config that actually works)

Model Context Protocol turns Claude Code from a code generator into something that can read your repo,
query a database, drive a browser and file a PR. The catch is that most "MCP setup" posts stop at
filesystem and never mention the traps. Here's a config I actually run, plus the gotchas.

A .mcp.json that earns its place

Drop this in your repo root; Claude Code auto-detects it and asks you to trust it once.

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

(Full pack with a postgres server and per-tool notes:
https://github.com/iskandaryv/anthropic-gateway-examples/tree/main/mcp)

The gotchas that cost real time

• The trust prompt is a feature. Claude Code refuses to load project MCP servers until you approve them — per project. That's the security boundary, not a bug.
• Give postgres a read-only role. An MCP server runs whatever SQL the model emits. A read-only role makes a hallucinated DROP physically impossible.
• Scope the GitHub token. repo + read:org is enough. Don't hand it an all-scopes classic PAT.
• MCP output is context. A 200-row query result or a full page fetch gets re-sent on the next turn. Big tool outputs inflate token usage fast — the thing that makes agentic sessions expensive.

It works the same through a gateway

MCP wiring is independent of where model calls go. Whether Claude Code talks to the official API or an
Anthropic-compatible gateway, the servers
are configured identically — the gateway only changes ANTHROPIC_BASE_URL, not .mcp.json.

Takeaways

• Start with filesystem + git + github; add browser/db only when a task needs them.
• Treat every server as an untrusted executor — least-privilege tokens, read-only DB roles.
• Watch token usage: tool results are the hidden cost of an agentic loop.

---

Disclosure: I run ClaudeAPIKey.dev, an Anthropic-compatible API
gateway — mentioned above only because MCP config is unaffected by it. "Claude" is a trademark of
Anthropic; we are not affiliated with Anthropic.