Skip to main content

Where your context window goes before you’ve said anything

Before a context window holds any of the work you care about, instructions, tool definitions, history, room for the reply, and a margin before compaction have already claimed their share. Here are five common Claude Code setups side by side, and what each leaves free.

Five setups, one window each

The biggest gap you can close with one setting: turning tool search on takes the MCP-heavy setup from 603K free to 775K.

  • Instructions
  • Tools and MCP
  • History
  • Reserved output
  • Compaction margin
  • Free for your work
  1. Lean CLI session

    810K free of 1M (81%)

    About the best case on a 1M model. The largest draw is the margin, which is a choice rather than a cost.
  2. MCP-heavy, tool search off

    603K free of 1M (60%)

    Tool schemas alone take 180K, which is more than instructions and history combined. All of it is spent before the model sees any evidence.
  3. …the same, with tool search on

    775K free of 1M (78%)

    Returns 172K tokens, or 17% of the window, for a configuration change rather than a discipline change. This is usually the largest single win available.
  4. Deep into a long session

    173K free of 1M (17%)

    History dominates everything, which is the shape that makes compacting worth it.
  5. A 200K window

    24K free of 200K (12%)

    Fixed overheads that look trivial against 1M are decisive at 200K.

Add your own

What you paste is read in your browser and never sent anywhere.

Which of these you can actually move

The segments look alike on a bar, but they are not equally movable. In rough order of how much they give back for how little effort:

  • Tools and MCP: turn tool search on. With tool search, Claude Code loads a tool’s schema when it needs one instead of loading every schema up front. In the MCP-heavy example that returns 172K tokens, 17% of the window, for a configuration change rather than a change in how you work. Loading fewer MCP servers helps too, but tool search is usually the largest single win available.
  • History grows on its own. Compacting replaces it with a summary, and subagents keep large file reads and tool results out of it in the first place.
  • Instructions move when you trim them. Instruction files, memory, and the skill listing are paid on every request, so a long one costs you twice: in room and in attention.
  • Reserved output follows from what you ask for. A longer answer needs more room. It is not really a lever.
  • The compaction margin is a choice. It is set by the autocompact threshold. Shrinking it buys room at the price of less slack before compaction fires.

This is accounting, not a performance model. What’s left is not uniformly useful: research on position sensitivity and on degradation as input grows points the same way, even if neither sets a universal threshold. A large free segment is permission to add material, not a promise that adding it helps.