brandonwie.dev
EN / KR
On this page
devops devopsclaude-codeai-ml

Anthropic MCP Context Budget Optimization

Anthropic-hosted MCP integrations consume ~71K tokens of your context window at session start — even when you never call them. Here is how to reclaim that budget.

• Updated April 6, 2026 • 3 min read

I was halfway through a complex refactoring session when Claude started losing track of files I had just read. The context window was running out, and I could not figure out why — until I ran /doctor and saw the numbers.

The Hidden Cost of MCP Integrations

Claude Code’s Anthropic-hosted MCP integrations — Gmail, Google Calendar, Notion, Sentry, Slack — load their full tool schemas into your context window the moment a session starts. Every tool definition, every parameter description, every type annotation gets serialized into tokens. Whether you call those tools or not, the space is consumed.

Here is what I found when I measured them:

Sentry:           ~15K tokens
Notion:           ~15K tokens
Calendar:         ~13K tokens
Slack:            ~8K tokens
Chrome-in-Chrome: ~6K tokens
────────────────────────────
Total:            ~57K tokens (before cleanup)

That is roughly 7% of a 1M context window spent on tool definitions alone. For shorter sessions or smaller models with tighter windows, this overhead is even more significant.

The Fix: Disconnect by Default, Enable on Demand

The solution is straightforward — disconnect all Anthropic-hosted integrations you do not actively use, and re-enable them only when needed:

  1. Run /mcp in Claude Code
  2. Disconnect integrations you are not using this session
  3. When you need Sentry for debugging: /mcp → enable Sentry → use it
  4. After you are done: optionally disconnect to reclaim context

Re-enabling is instant. OAuth reconnects transparently — no re-authentication needed.

Before and After

Before: ~71,834 tokens consumed by MCP tools
After:  ~14,000 tokens (only context7, playwright, chrome-devtools remain)
Saved:  ~57,000 tokens (~5.7% of 1M context window)

When This Matters

This optimization is most valuable during:

  • Long development sessions where you are reading large files and accumulating context
  • Large codebases that already consume significant context just from file reads
  • Complex multi-step tasks where Claude needs to hold many details in memory

It matters less for quick sessions or when you are specifically doing Sentry triage or Slack-heavy communication workflows. If you frequently toggle integrations on and off, the reconnection friction may outweigh the savings.

Key Takeaway

Context window space is a finite resource. Tool schemas are loaded eagerly, not lazily — they cost tokens whether you use them or not. Audit your MCP integrations with /doctor or /mcp, disconnect what you do not need for the current session, and reclaim that budget for actual work.

Comments

enko