The context-window tax

Every MCP server you connect advertises its tools by dumping their full JSON schemas into the model's context on every turn — whether the agent uses them or not. Connect enough servers and a meaningful slice of the window is gone before the model reads a single line of your actual problem. We call it the context-window tax.

The method

We measured a representative three-server setup — github (20 tools), supabase (15), playwright (10): 45 tools total — serializing each server's tools/list payload before and after mcpfold's per-tool curation. The tokenizer is the widely-cited approximation of 1 token ≈ 4 characters of JSON. Exact counts vary by model, but the relative reduction is stable because both sides are measured identically.

The result

Curating down to the 9 tools actually needed cuts tool-schema tokens by ~80% (7,476 → 1,497). No extra configuration — the shim already in the launch path does the filtering, so the win is free.

Curation turns "connect every server" from a context-window tax into a cheap, fast, focused toolset.

Reproduce it yourself with pnpm --filter @mcpfold/proxy bench, or try the interactive calculator on the homepage. Full methodology lives in the benchmark docs.