MCP Tool Search: How Claude Code Fixed Context Window Bloat

One of the biggest problems right now for everyone using AI coding tools is context window and MCP.

We only have 200K context window. And we don't even get full allocation for that when we code. Between system prompts, conversation history, and tool definitions, you're already starting at a deficit.

Well, Claude Code just released MCP Tool Search.

The Context Tax

The biggest issue with MCPs is they're powerful but expensive.

Just 5-6 MCP integrations will fill up your context window and you can't even do anything. You connect Figma, Playwright, GitHub, a couple custom servers, and suddenly 30-50% of your context is gone before you type a single prompt.

Opus is powerful. But what's the point of using it when you're running out of session every 10-15 minutes?

We can't ignore MCPs either since it's the easiest and most popular way to extend Claude's capabilities. Browser automation, database access, file operations, custom APIs. This is what makes Claude Code actually useful for real work.

So we were stuck. Use MCPs and burn context. Or skip MCPs and lose functionality.

What We Tried

A couple months ago, solutions started coming out. Anthropic, Cloudflare, Docker all mentioned their suggestions and approaches. People tried them, including me.

• Code execution wrappers

• Skills with lazy loading

• Universal MCP configs

• Manual tool prioritization

They all have pros and cons. Some broke tool discovery. Others added latency. None felt like real solutions.

I've been waiting for an official fix. And now it's here.

The Fix

It's awesome and intuitive.

Connect whatever MCP you want and Claude will search and load relevant tools for you. Instead of preloading every single tool definition at session start, it searches on-demand.

Here's how it works:

1. Threshold detection: If your MCP tools exceed 10% of context, Tool Search activates automatically.

2. Semantic search: When you need a tool, Claude searches for it by keyword and similarity. Need GitHub tools? It searches "github" and loads just those definitions.

3. On-demand loading: Tools only consume context when you actually use them. Switch tasks? Context shifts with you.

No config needed. It just works.

The Proof

Before: MCP tools eating 20-50% of context window

After: MCP tools loaded on-demand, effectively 0% until needed

Simon Willison said it best: "Context pollution is why I rarely used MCP, now there's no reason not to hook up dozens or hundreds of MCPs."

That's exactly right.

If It's Not Working Yet

Since it's rolling out, it might not be available for you even though your CLI is up to date.

For weird reasons, it happened to me. If it's happening to you, do this:

source ~/.zshrc
export ENABLE_TOOL_SEARCH=true

That's it.

Check with /context. If MCP tools shows "loaded on-demand" instead of a token count, it's working.

Call some MCP tool and you'll see MCPSearch appear when Claude finds the right integration.

Go Wild

Now I'm enabling all the MCPs I've been waiting to use. I'm Lovin' It.

Notion, Linear, Exa, Vercel, database tools, the whole stack. There's no penalty anymore.

The barrier that kept most people at 2-4 servers is gone. Connect everything. Let Claude figure out what it needs.

Time for you to get wild too. And let's find something else to complain to Anthropic about lol