Like many Claude Code users, I went on an MCP (Model Context Protocol) installation spree.
Every time I found an interesting MCP server, I added it. GitHub, browser automation, databases, AWS… before long my Claude Code had become a Swiss Army knife.
One day I started a brand-new Claude Code session and simply typed:
Hi
Out of curiosity, I checked the context usage.
To my surprise, the initial context was already around 150,000 tokens, even though I’d only typed a single word.
That also meant the session had already consumed more than $1 worth of tokens before doing any useful work.
More importantly, every conversation was carrying a huge amount of MCP tool definitions that I wasn’t actually using. Besides the higher cost, a significant portion of Claude’s context window was already occupied before the real work even began.
Finding the cause
After a bit of investigation, I realized the issue wasn’t Claude Code itself.
Every installed MCP server exposes its available tools to Claude. As I added more and more MCP servers, all of those tool definitions were included in the initial context for every new session.
The more MCP servers I installed, the larger — and more expensive — the starting context became.
The one-line fix
While reading the Claude Code documentation, I discovered this environment variable:
export ENABLE_TOOL_SEARCH=trueI enabled it without uninstalling or disabling a single MCP server.
The difference was immediate.
| Metric | Before | After |
|---|---|---|
| Initial context | ~150k tokens | ~30k tokens |
| Startup cost | >$1 | ~$0.06 |
| Tool loading | Every conversation loads all MCP tool definitions | Only relevant tools are loaded |
| Context availability | Large amount occupied before work begins | Most context remains available for actual tasks |
That’s roughly an 80% reduction in initial context size and around 20× lower startup cost.
Why it works
By default, Claude Code loads information about all available MCP tools into the conversation context.
When ENABLE_TOOL_SEARCH=true is enabled, Claude Code changes its strategy.
Instead of loading every tool definition up front, it first performs a lightweight search to determine which tools are relevant to your request. Only those relevant tools are then added to the context.
Conceptually, it changes from this:
Your prompt
+ Tool A
+ Tool B
+ Tool C
+ Tool D
...
+ Tool Zto this:
Your prompt
+ Search available tools
↓
Only Tool C
Only Tool GAs your collection of MCP servers grows, this optimization becomes increasingly valuable.
Who should enable this?
If you only have a couple of MCP servers installed, the impact may be minimal.
But if you’ve accumulated dozens of MCP servers like I have, this is an easy optimization that lets you keep all of them without bloating every conversation’s context.
Final thoughts
The biggest surprise wasn’t that saying “Hi” cost over a dollar.
It was discovering that every new Claude Code session started with around 150,000 tokens of context — most of it from MCP tool definitions I wasn’t using.
After enabling ENABLE_TOOL_SEARCH, the initial context dropped to under 30,000 tokens, the startup cost fell from over $1 to about $0.06, and much more of the context window was available for the work I actually wanted Claude to do.
If you’re an enthusiastic MCP collector, this is one setting that’s absolutely worth enabling.
Reference: Scale with MCP tool search