Evidence-BasedMulti-DimensionalReviewable ReportsFree
Code paths, network calls, dependency risk — deliverable verdict in minutes.
High-Risk Skills Identified
View full risk board →Not a popularity chart. These are skills flagged for blocking or manual review, ranked by risk score.
| # | Skill | Verdict | Risk |
|---|---|---|---|
| 01 | math-calculator | Block | 100 |
| 02 | messenger_send_node | Block | 95 |
| 03 | vnstock-env-setup | Block | 92 |
| 04 | luci-memory | Block | 85 |
| 05 | video-content-pipeline | High Risk | 78 |
| 06 | mcp-server-builder | High Risk | 75 |
| 07 | xhs-note-analyst | High Risk | 75 |
| 08 | birth-system-manager | High Risk | 75 |
View full risk board 571 high-risk samples identified
Trust Criteria
How we decide whether a skill deserves trust
Not from demos, not from download counts, but from the evidence left behind in code, metadata, and runtime intent.
Declared vs actual capability
We line up claimed resources against the real shell, network, filesystem, and environment behavior inferred from the code.
Hidden execution and egress
We check for encoded execution, command composition, outbound URLs, hard-coded IPs, credential patterns, and dangerous command chains.
Supply chain and dependency hygiene
We look for unpinned packages, known vulnerabilities, suspicious download paths, and third-party components that widen the risk surface.
Reviewable evidence
Every judgment should land on files, lines, artifacts, dependencies, and attack-chain steps instead of ending at a score.
Live Signals
Recently reviewed skills
These samples show the issues the system is seeing in real submissions. The real user question is not what a skill does, but whether it deserves installation.
Live Signals
Report Contract
01
What does the skill claim to do, and what does the code actually do?
02
Does it hide execution, egress, credential activity, or an identifiable attack chain?
03
Which files, dependencies, or artifacts are the key reasons it should not be trusted?
04
If the team keeps using it, should they block, isolate, or manually review it next?
Decision Flow
From input to install decision in four steps
The goal is not to keep users on the scan page. It is to move them to an actionable decision as fast as possible.
Receive input
Repos, skill pages, archives, and raw URLs enter the same review pipeline.
Extract evidence
We first read the file tree, sensitive files, artifacts, dependencies, and declared metadata.
Infer intent
We then compare declared capability with actual behavior to see whether the skill drifts, deceives, or attacks.
Make the decision
The report ends in a block, review, or allow recommendation with evidence attached.
Next Action
Throw the skill in first. Decide whether to trust it after.
If you already have a target to review, start now. The output is structured for engineering, security, and audit teams to act on together.
No account required · Free for open repos