Scan Agent Skills and MCP configs for risky commands, prompt injection, secret access, and obfuscation. Check GitHub repos, files, folders, or ZIPs locally.
Local static analysisNo signup · no skill content uploaded
What would you like to scan?
Each method is independent
Review before installing
Use a public GitHub repository or a direct HTTP(S) file URL. Loading starts only when you check.
Local browser scan by default; optional online checks run only when you select them.
About this tool
Made for the task at hand.
Review Agent Skills and MCP configurations before giving them access to your tools, files, or credentials. KitVerse combines its own configurable rules with validated NVIDIA SkillSpector static patterns to detect prompt injection, hidden or encoded instructions, dangerous commands, secret access, persistence, MCP tool poisoning, and suspicious dependencies. Scan a GitHub repository or directory, local files and folders, ZIP packages, or pasted text; then review file-and-line evidence, OWASP Agentic and MITRE ATLAS mappings, and export JSON, Markdown, or SARIF 2.1.0. Static findings guide manual review and never prove that a package is safe.
How to use
Choose Link, File, or Text. Enter a public GitHub repository, directory, or direct file URL; select local files, a folder, or ZIP; or paste SKILL.md, MCP configuration, or script content.
Select Check Now. When multiple skills are discovered, scan one card or use Scan All Skills with independent progress and cancellation; list filters never change the batch scope.
Review the score, verdict, source evidence, recommendations, and unscanned entries. Open Security rules to inspect or configure future scans, and explicitly choose OSV.dev or Koi Clawdex only if you need those optional online checks.
Copy or export the individual or aggregate report as JSON, Markdown, or SARIF 2.1.0. Investigate possible false positives and test unknown packages in an isolated environment before installation.
Your work stays yours. Tool inputs are processed locally in your browser and are never sent to KitVerse servers. Clear the workspace when using a shared device.
Common questionsWhat skills and packages can I scan?
Scan a public GitHub repository or tree directory, a direct SKILL.md or ZIP URL, local files and folders, ZIP packages, or pasted text. Repository, folder, and archive inputs first list discovered SKILL.md files for individual or batch checks. Markdown, scripts, YAML, JSON, MCP configurations, and common dependency manifests are supported; unsupported entries are identified as unscanned.
What do the risk score and verdict mean?
Severity-weighted findings produce a score capped at 100: 0–20 Safe, 21–50 Caution, 51–80 High risk, and 81–100 Critical. The score prioritizes manual review; it does not measure certainty or prove that a skill is safe or malicious.
Does the scanner upload my skill or use AI?
No AI model or KitVerse server analyzes your content. Default static scanning runs locally in a browser Worker. A remote URL is fetched directly from its host only after you submit it. Optional OSV.dev and Koi Clawdex queries require separate clicks and send only dependency coordinates or a skill name—never skill text, scripts, or report contents.
Can a clean result prove an Agent Skill is safe?
No. Static rules and limited payload decoding can miss dynamically constructed behavior and can flag benign examples. Treat every report as review guidance, inspect the evidence and false-positive notes, and validate unknown packages in a sandbox.
Can I configure rules and export audit reports?
Yes. Search, inspect, enable, or disable local and validated NVIDIA SkillSpector rules; changes are stored only in this browser and apply to future scans. Upstream updates are committed as a validated static rule pack by the data workflow, never fetched during a browser scan. Export individual or aggregate results as JSON with file hashes, readable Markdown, or SARIF 2.1.0 for compatible review tools. Local scanning and exports remain available offline; remote loading and optional intelligence checks do not.