A recent analysis of 25 popular Model Context Protocol (MCP) servers revealed a significant, often undisclosed, "context tax" imposed by tool schemas. These schemas are injected into every model request, consuming valuable context window space and increasing costs. The study found a 544x variation in this overhead, ranging from 35 to over 19,000 tokens per request, with some combinations like notion-mcp and github exceeding 23,000 tokens before any user input. The analysis also identified security concerns, with many servers reading secret environment variables and some exhibiting network egress to unexpected domains or bundling minified code that obscures its network activity. AI
IMPACT Highlights hidden costs and potential security risks in AI agent infrastructure, urging developers to measure and manage token usage.
RANK_REASON Analysis of a technical protocol and its implementation costs. [lever_c_demoted from research: ic=1 ai=0.7]
- api.agnost.ai
- api.exa.ai
- CLIENT_IP_ENCRYPTION_KEY
- exa
- github
- login.microsoftonline.com
- MCP
- Model Context Protocol
- notion-mcp
- OPENAI_APPS_CHALLENGE_TOKEN
- upstash-context7-mcp
- wartzarbee
- wartzar-bee/ci-guardrail
- @wartzar-bee/tokenscope
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →