A Measurement Study of Model Context Protocol Ecosystem

作者：Hechuan Guo, Hao, Yongle, Yue Zhang, Minghui Xu, Penghao Lv, Jiezhi Chen, Xiuzhen Cheng · 发表于：arXiv (Cornell University) · 年份：2025 · DOI：10.48550/arxiv.2509.25292 · 被引用次数：2 · 研究领域：Software System Performance and Reliability、Topic Modeling、ICT in Developing Communities

The Model Context Protocol (MCP) ecosystem now exceeds 14,800 publicly addressable servers, but reliability claims about individual servers are typically issued as point-in-time snapshots or static manifest scores. This paper describes the continuous behavioral telemetry methodology operated by Dominion Observatory since 2026-04-08, in which (a) an active probe fleet pings registered MCP endpoints on a fixed interval, (b) agent-reported interaction outcomes are recorded under an explicit provenance schema, and (c) the resulting time-series is exposed as a CTEF v0.3.2 §4.5 evidence_provider with a re-runnable conformance harness anchored in git history. Across 38 days of operation we report 88,216 interaction records over 14,820 tracked servers, with self-attested conformance vectors published at stable workers.dev endpoints. The methodology is differentiated from existing MCP empirical studies (which are snapshot-based) by three properties: continuity, commit-anchored proof, and cross-validation across four independent specs (CTEF v0.3.2 §4.5, x402 Trust-Provider Interface v0.1 PR #2300, A2A evidence_ref v1.0, ERC-8004 endpoint-health attestation v1.0). All data substrates, schemas, and verification scripts are publicly fetchable on free-tier infrastructure.