Benchmarking MCP Server Context Overhead
Crazyblind · reddit · 2026-07-11
The author open-sourced a CLI, lap, to measure the context window tax on agent-facing interfaces, covering MCP servers, OpenAPI menus, and local tool stacks.
Core Findings
- Among 20 public MCP servers, the Notion official server consumes 21,411 tokens just for tool definitions in a single session.
- Firecrawl takes about 18,511 tokens.
- Connecting all 20 servers eats up roughly 64k tokens before startup.
- Despite having only one tool, sequential-thinking requires 921 tokens due to its lengthy description.
Further Experiments
- Compressing the same tools into tighter signatures drops the total cost to 3,100 tokens, a roughly -95% reduction.
- lap stack reads your Claude Desktop / Cursor configurations to calculate startup costs.
- lap lint --mcp "<command>" detects missing tool descriptions, undocumented parameters, overly long definitions, and missing required fields.
Results & Impact
- The author launched a monthly-updated public API leaderboard showing that directly converting OpenAPI to MCP menus accounts for 11.2M tokens in scale, meaning 82% of the cost appears recoverable.
- They also tested the "layered schema" proposal from the spec, finding that the discovery tier saves an average of 87%.
- The project is MIT licensed with no commercial wrapping; the author notes it typically spots issues within ten seconds, and upstream maintainers usually fix them quickly.
Related event: MCP Tool Schema Overload Impairs Agent Accuracy(3 posts)→
More from coding & agent
- Hugging Face wraps 6-month Training Agents series with open recordings and code — SergioPaniego · 2026-10-26
- Agent verified 'success' after double-voting: a lesson in harness vs model bugs — TejasKumar_ · 2026-10-26
- diginomica: agentic automation hype is overdone, and 'The Hackening' of enterprise security has begun — jonerp · 2026-10-26
- 'Vibe verification': ~60% of AI-written formal specs prove the wrong property — basedjensen · 2026-10-26
- Devin launches Security suite powered by Claude Mythos with $100M+ credits — DevinAI · 2026-10-26
- HackBench injects reward hacks into agent trajectories to stress-test monitors — DanielKhashabi · 2026-10-26