When Your AI Has Too Many Tools

AI assistants are getting more powerful. We can now connect them to multiple services through Model Context Protocol (MCP) servers—giving them access to Gmail, GitHub, file systems, and more. But there's a hidden cost: every available tool gets loaded into your context window, even if you don't use them all.

I experienced this firsthand with Claude. I had configured 11 MCP servers and 128 tools. But when I audited what Claude could actually see, only 5-6 servers were working properly. The rest? Dead weight in my prompt, consuming tokens I could've used for actual work.

The Token Tax of Tool Overload

Each tool takes space in your prompt. Let me show you what I had:

That's ~27,000 tokens spent on tools alone. Many of them were redundant or barely used. The problem: Claude would see all these tools, parse them all, and consider using them—even when simpler, built-in alternatives existed.

The Aggressive Optimization

I didn't just cut back. I went aggressive. Here's my final stack:

Removed:

• Filesystem – Desktop Commander does everything better
• Windows CLI – Redundant with Desktop Commander
• MCP Installer – I don't install packages during chats
• Brave Search – Using built-in web_search instead
• Sequential Thinking – Native reasoning works fine

Kept:

• Gmail – Unique capability, irreplaceable
• GitHub – Version control and code operations
• Desktop Commander – File system + CLI combined
• Memory Metadata – My personal memory system
• Puppeteer – Browser automation for specific tasks
• Vercel – Deployment automation

From 11 servers → 6 servers From 128 tools → 89 tools From ~27k tokens → ~17k tokens

That's a 37% context window improvement.

Why This Matters

Context window is currency in AI. Every tool you don't actually use is a theft of that currency. Before my optimization:

• Prompt was bloated
• Claude had to parse irrelevant tools
• Decisions took longer (cognitive overhead)
• Token waste accumulated across every conversation

After:

• Lean stack with clear purpose
• Faster tool discovery
• More tokens for actual work
• Philosophy: Essential only

The MCP Discovery Pattern

Here's what I learned: Most MCP servers are "nice to have," not "need to have." When evaluating tools, ask:

1. Is this better than what's built-in? (Filesystem is not better than Desktop Commander)
2. Do I actually use this? (I never installed MCP packages mid-chat)
3. Does this reduce cognitive overhead or add it? (Too many search tools add overhead)

What I'd Recommend

If you're building your own MCP stack:

• Start with 3-5 core tools (email, code, files, research)
• Add specialized tools only when you can measure the value
• Audit quarterly – remove what you're not using
• Remember: A focused toolkit beats a comprehensive one

Your AI is most effective when it has exactly the tools it needs, not everything it could possibly want.