skill-bouncer

TSA for agent skills. Scan the skill before install, deny overreach with a readable receipt, then bundle the safe ones.

Skill Bouncer is an open-source Next.js app for reviewing third-party SKILL.md files before they reach an agent bundle. It parses the skill, infers implied capabilities, applies deterministic policy rules, shows an explainable ALLOW / REVIEW / DENY receipt, and lets you compose a bundle from the approved set.

Live demo: skill-bouncer.vercel.app

Additional screenshots:

• Home / intake state
• Seeded research dataset
• Denied receipt with quoted evidence
• Approved bundle summary

Most agent ecosystems still treat skill installation as a trust problem. A skill can claim a narrow job while quietly asking for shell access, broad filesystem reads, or secret access. Skill Bouncer is intentionally narrower than a full runtime AI firewall: it is a fast intake gate for pre-install review.

That positioning matters:

• Runtime AI security platforms focus on live enforcement and governance.
• Eval tools focus on generic testing and prompt attacks.
• MCP registries focus on discovery and distribution.
• Skill Bouncer sits earlier in the workflow: *before* a third-party skill is trusted or bundled.

• Upload your own markdown skill batch anonymously
• Load the built-in research-10 dataset with curated allow/review/deny ground truth
• Inspect receipts with exact quoted evidence lines from the skill
• See provenance links for seeded research samples
• Build a clean bundle from approved skills only

Included in v1:

• Plain SKILL.md intake
• Deterministic capability extraction and poli