MCP servers eat 29k tokens before you type a word — dev ships mcp-tax CLI to measure the tax
princejain756 · reddit · 2026-10-12
The author connected to all 9 MCP servers in his configs as a plain client and counted the tokens their tool/prompt/resource schemas inject into context: claude-code 14.3k (Bash schema alone 2.5k), chrome-devtools 5.3k, blender 3.5k—29k total, or 90% of a 32k window, explaining why local models feel lobotomized with many MCP servers. Official reference servers are lean (fetch is just 235 tokens); description verbosity drives the cost. He packaged the measurement as a CLI that auto-discovers configs, spawns servers over stdio, counts with o200k, and reports per-server totals as % of 200k/128k/32k windows: npx github:princejain756/mcp-tax.
More from coding & agent
- 10 open-source projects extending AI from chatbots to docs, browsers and memory — Shruti_0810 · 2026-10-12
- LangChain founder lays out three models for enterprise agent identity — hwchase17 · 2026-10-12
- jax-graft: an AI-built JAX backend runs JAX on Apple Silicon GPUs — twiecki · 2026-10-12
- Ditch shadcn and Tailwind patterns to avoid the AI-slop web look — michalmalewicz · 2026-10-12
- Student struggles with agentic coding: Claude Code specs + Antigravity still miss details — 3ATAE · 2026-10-12
- All of science embedded and free: 200M papers searchable by AI agents, no API key — pbaylies · 2026-10-12