MCP (Model Context Protocol) servers can make or break your agent performance. This guide helps you avoid common pitfalls.
Every MCP server you install consumes context tokens before you even ask a question.
200K advertised context window
- 16K model overhead
- 16K harness overhead
= 176K USABLE
With typical MCPs installed:
- GitHub MCP: ~4K tokens
- Chrome MCP: ~5K tokens
- AWS Lambda MCP: ~40K+ tokens (!)
= Down to <100K usable
Real impact: With heavy MCPs loaded, you lose 28%+ of your context before doing any work.
"Do not install the GitHub MCP server. Just prompt the LLM to use the GitHub CLI." - Geoffrey Huntley
| Task | MCP Approach | CLI Approach | Winner |
|---|---|---|---|
| GitHub issues | GitHub MCP (~4K overhead) | gh issue list |
CLI |
| Git operations | Git MCP | git status |
CLI |
| File search | Custom MCP | rg, fd |
CLI |
| API calls | Custom MCP | curl |
CLI |
Why CLI wins:
- Zero token overhead
- Claude already knows these tools
- No schema bloat
Run /context in Claude Code to see token breakdown:
System tools: 16.7k tokens (8.4%) ← Built-in
MCP tools: 45.0k tokens (22.5%) ← Problem area
Look for mcp__ prefixed entries to identify culprits.
Claude Desktop:
Windows: C:/Users/[USER]/AppData/Roaming/Claude/claude_desktop_config.json
Mac: ~/Library/Application Support/Claude/claude_desktop_config.json
Claude Code:
~/.claude.json (user config)
.mcp.json (project config)
Start with no MCPs, add only what you need:
{
"mcpServers": {}
}Then add only project-specific servers via .mcp.json in your project root.
| MCP | Token Cost | Alternative |
|---|---|---|
| GitHub MCP | ~4K | Use gh CLI |
| AWS Lambda MCP | ~40K+ | Use aws CLI |
| Chrome MCP | ~5K | Disable when not needed |
| Memory MCP | ~2K | Usually unnecessary |
In ~/.claude.json:
{
"claudeInChromeDefaultEnabled": false
}Just delete it from config. Use gh CLI instead:
gh issue list
gh pr list
gh pr createKeep multiple configs:
claude_desktop_config.minimal.json # Daily use
claude_desktop_config.aws.json # When debugging AWS
claude_desktop_config.full.json # Rare, specific tasks- Start minimal - Empty MCP config by default
- Audit regularly - Run
/contextto check usage - Prefer CLI -
gh,aws,kubectlover MCP equivalents - Load on demand - Switch configs for specific tasks
- Project-specific - Use
.mcp.jsonfor project needs
For advanced MCP optimization including custom server development and token budgeting:
- Premium Guides: Complete AI Development System — Includes the full MCP Development Guide