Governance for agent instructions: audit, score, and pin MCP servers, Agent Skills, and prompts. Deterministic; findings carry a reference and a fix.
View on GitHub ↗Tool descriptions and skill files are injected into a model's context on every request. They decide which tool gets called, with what arguments, whether the client asks before something is destroyed, and whether a skill fires at all. They are production configuration — and almost nobody reviews them, versions them, or notices when they change.
sounding audits them, scores them with a formula you can check, writes the corrections, and pins them so you find out when they drift.
------ MCP servers permission, blast radius, injection surface, transport, secrets Agent Skills triggering, progressive disclosure, dangerous commands, portability Prompts output contract, failure behaviour, interpolation safety, contradictions
They are the same problem. A skill and a tool description are both text handed to a model, living in a repo, versioned, drifting — so they share one engine and one report.
Every finding carries a reference. Not "this looks risky" — the spec section or guidance that says why. A finding without a source is an opinion, and this tool does not ship opinions.
Every finding carries a fix, or a question. When the correction is unambiguous, it says what to change. When it depends on intent the tool cannot infer, it asks — with options, capped at three per run:
And then it writes them. Answers become edits to your descriptor, shown as a diff before anything is touched:
Anything generated that still needs a human is marked TODO in the output. The tool does not write prose and present it as finished — and in a SKILL.md it rewrites the frontmatter only. Body text is y
From the project README.
Add the radar badge to your README — it shows your project was picked up by MCP Radar and links to this page:
[](https://mcp.liqiwa.com/s/alinotfoundbtw--sounding.html)
Find which APIs of your pinned Python dependencies changed after your coding model's training cutoff, and give the agent short AGENTS.md / CLAUDE.md notes from a static API diff. No model calls, no AP
eltociear/mcp-auditScan MCP servers & AI-agent skills for malicious patterns (prompt injection in tool descriptions, credential exfil, download-and-execute). 17 patterns / 60 signatures, zero-dep. Measured over 196 publ
valorisa/Reverse-Engineer-AnythingGuide pédagogique non officiel en français sur REA (Reverse Engineer Anything) : analyser des applications et des binaires avec un agent de code via MCP et CLI, Hopper ou Ghidra, preuves à l'appui. In
HUA503/mcprismFast single-binary security scanner for MCP (Model Context Protocol) servers — config audit, live capability enumeration, tool-poisoning detection & risk grading. Zero deps, fully offline — trivy for
1105216375-alt/resimindNeuro-symbolic agents with verified memory. MCP math tools, an algebra repair Skill, and an open challenge. Verify solutions, learn checked rules, and reuse.
sepehr071/awesome-iran-mcp19 open-source MCP servers (308 tools) that give AI agents the Iranian web: Digikala, Snappfood, Jabama, MrBilit, Paziresh24, Taaghche and more. Prices in Toman.
The top new MCP servers of the week, every Monday. No spam, unsubscribe anytime.