Context & MCP Tradeoffs

MCP servers are powerful, but they come with hidden costs. Understanding these tradeoffs will help you use them effectively.

The Problem: Context Rot

Every piece of information in Claude’s context window competes for attention. Research shows that LLM performance degrades as context fills up—a phenomenon called context rot.

Key Finding

At 32k tokens, 11 out of 12 tested models dropped below 50% of their performance in short contexts. The degradation is worse for complex tasks requiring multi-step reasoning.

What the Research Shows

• Performance degrades before you hit context limits
• Complex reasoning tasks suffer most
• More context ≠ better answers
• There’s an “optimal context window” much smaller than the maximum

How MCPs Affect Context

MCPs add context in two ways:

1. Tool Definitions

Each active MCP adds its tool definitions to Claude’s context:

MCP: context7
├── search_docs tool definition (~500 tokens)
├── get_documentation tool definition (~400 tokens)
└── list_libraries tool definition (~300 tokens)

Total: ~1,200 tokens per MCP (just for definitions!)

2. Tool Responses

When Claude uses an MCP, the response goes into context:

> use context7 to get FastAPI docs

[MCP Response: 8,000 tokens of documentation]

That documentation stays in context for the entire conversation.

The Math

Component	Tokens
Your CLAUDE.md	500-2,000
System prompt	~3,000
Each MCP (definitions)	500-1,500
Each MCP response	2,000-20,000
Your conversation	Variable

With 5 MCPs active and a few queries:

• 5 MCPs × 1,000 tokens = 5,000 tokens
• 3 MCP responses × 8,000 = 24,000 tokens
• System + CLAUDE.md = 4,000 tokens
• Overhead: 33,000 tokens before you ask a coding question

The 40% Rule

Best Practice

Keep your context usage below 40% of the maximum for optimal performance. Above 70%, expect noticeable degradation.

Why 40%?

• Leaves room for Claude’s reasoning
• Maintains attention on recent/relevant information
• Prevents “lost in the middle” effects (Claude forgetting middle context)

How to Check

> /context

Or watch for warning signs:

• Claude forgets earlier instructions
• Responses become less coherent
• Claude asks questions you already answered

Mitigation Strategies

1. Minimize Active MCPs

Only enable MCPs you’re actively using:

# Morning: Working on frontend
claude mcp enable playwright
claude mcp disable postgres

# Afternoon: Database work
claude mcp enable postgres
claude mcp disable playwright

2. Use MCPs Strategically

Don’t fetch documentation you don’t need:

# Bad: Fetches entire FastAPI docs
> use context7 to get FastAPI documentation

# Good: Fetches only what you need
> use context7 to find FastAPI's dependency injection syntax for database sessions

3. Start Fresh Sessions

After heavy MCP usage:

# Clear and restart
> /clear

# Or start a new session
exit
claude

4. Use /compact Proactively

> /compact

This summarizes your conversation, including MCP responses, into a smaller context.

5. Be Explicit About What You Need

# Bad: Returns everything
> query the database for user statistics

# Good: Returns minimal data
> query: SELECT COUNT(*) as total, status FROM users GROUP BY status

Monitoring Context Health

Signs of Context Rot

1. Forgetting: Claude asks about things you already discussed
2. Inconsistency: Responses contradict earlier decisions
3. Confusion: Claude conflates different parts of the conversation
4. Degradation: Quality drops noticeably

Recovery

When you notice these signs:

# Option 1: Compact and continue
> /compact

# Option 2: Start fresh with context
> /clear
> Here's what we were working on: [brief summary]

# Option 3: New session for new task
exit
claude