AI
MCPGuard: Finding Conformance Gaps in MCP SDKs via Dual-Track Analysis and Dynamic Adjudication
Publication: cs.AI updates on arXiv.org
Author: Xikai Sun, Kebin Liu, Fan Dang, Yawen Zheng, Zihao Yang, Yunhao Liu
Source published: 04/11/2026, 05:00:00 UTC
Briefing created: 04/11/2026, 06:47:11 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 04/11/2026, 05:18:24 UTC.
Briefings
MCPGuard is an automated LLM-based framework designed to detect specification-to-SDK conformance gaps in Model Context Protocol (MCP) SDKs. It operates via two tracks: a structural track aligning JSON Schema constraints with normalized SDK models, and a behavioral track extracting normative prose rules and auditing them with LLMs. An evidence-aware dynamic adjudicator then inspects complete source context to build Proof of Concepts (PoCs) validating candidates. The framework was tested on official Python 1.26.0 and TypeScript 1.29.0 SDKs, confirming 13 and 19 findings respectively, none of which were exposed by existing black-box baselines. These gaps can systematically propagate downstream when not addressed.
Key points
- MCPGuard uses a dual-track approach combining structural alignment of JSON Schema constraints with normalized SDK models and behavioral auditing of normative prose rules.
- The framework identified 13 conformance gaps in the official Python 1.26.0 SDK and 19 gaps in the TypeScript 1.29.0 SDK, findings not detected by existing black-box conformance tools.
Context
The Model Context Protocol (MCP) enables large language model agents to connect to external tools and services through official SDKs. Existing conformance tools primarily test running endpoints, failing to reveal whether SDK models enforce specification requirements. MCPGuard fills this gap by automating the detection of specification-to-SDK conformance gaps.
Limitations
Evidence is limited to official Python 1.26.0 and TypeScript 1.29.0 SDKs; open MCP server testing showed gaps propagate downstream but specific propagation mechanisms were not fully characterized.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 04/11/2026, 06:47:11 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeibs4sznso5p7jeb4utrifvoz2cr6xqg4apy3hvuky6hnf62lqsv6e