01 · SCAN
Inspect before install
Analyze instructions, scripts, dependencies, and tool calls as one attack surface—not isolated files.
Public beta · Static pre-install security
Know what youragent can do.
Longxia scans source and instruction files from public GitHub agent skills before they touch your machine. The scanner never executes repository code.
No credit cardEvidence-backed findingsPublic scans stay free
Sample skill risk report
openclaw/research-agent · v2.4.1
Risk score
82/ 100
FILES24
RULES127
SCAN TIME4.8s
3 actionable findings
Hidden shell execution
scripts/install.sh · line 42
Unscoped credential access
SKILL.md · tool permissions
Remote content injection
references/setup.md · external URL
Illustrative preview. Open the real report to review source evidence from the Git commit actually scanned.
STATIC RULESET · SAMPLE DATAONLINE
Supported source and instruction formats
SKILL.mdAGENTS.mdpackage.jsonpyproject.tomlShell scriptsYAML + TOML
The missing trust layer
Agents install trust at machine speed.
A skill is more than code. It is instructions, dependencies, permissions, and live access to your files and accounts. A clean repository can still teach an agent to do something dangerous.
Instructions
Natural-language files can ask an agent to ignore boundaries or hide behavior.
Commands
Install scripts can execute remote payloads, read secrets, or modify broad filesystem locations.
Dependencies
Unpinned packages and downloads can change what eventually reaches the machine.
What the public beta does
Security that speaks agent.
Longxia combines instruction and source-pattern checks, then exposes the evidence. It is a narrow pre-install review tool, not runtime protection.
02 · MAP
See the real blast radius
Turn vague manifests into a plain-English map of files, secrets, services, and commands a skill can reach.
03 · TRACE
Review the evidence
See the rule, file, line, source excerpt, severity, and practical remediation behind every finding.
04 · SHARE
Share one point-in-time report
Send an unlisted report tied to the exact Git commit. Reports remain available for 30 days.
Simple by design
From unknown to understood.
A security decision your developer, IT lead, and founder can make from the same report.
01
Paste
Enter a public GitHub repository, folder, or file URL.
02
Analyze
Longxia evaluates code, natural-language instructions, and permissions together.
03
Review
Use the risk score, capability map, and source evidence to decide what to inspect next.
Zero execution by design
Fetched files are size-limited, parsed as text, and never run or installed by Longxia.
Real public scans
Three skills. Three different decisions.
Each case is a point-in-time static scan of a public GitHub source. Open the report to inspect its commit, files, rules, evidence, and remediation.
- BlockedRisk score 49 / 100
openai/skills · sentry
A remote installer should stop the install.
The instructions pipe a downloaded installer directly into Bash. Verify the source and checksum before proceeding.
- Findings
- 2
- Commit
- 49f948faa925
- ReviewRisk score 24 / 100
anthropics/skills · webapp-testing
Local command execution needs review.
Two Python calls start or run local commands. That may be intentional for testing, but the capability deserves human review.
- Findings
- 2
- Commit
- b29e7cf65e5c
- PassRisk score 0 / 100
vercel-labs/agent-skills · web-design-guidelines
No known high-risk pattern matched.
This one-file instruction skill triggered no current rules. A pass is a review starting point, never a safety guarantee.
- Findings
- 0
- Commit
- 7c180d9044c9
Results describe exact commits and the Longxia ruleset at scan time. A verdict is not a malware classification or endorsement; project names identify public sources only.
Public beta access
Start with one public skill.
The implemented scanner is free while Longxia validates the workflow with real users. There is no checkout or paid plan today.
Available now
Public scanner
For reviewing public GitHub agent skills before installation.
- 10 anonymous scans per day
- Repository, folder, and file URLs
- Capability map and source evidence
- Shareable reports available for 30 days
Roadmap · Not yet available
Private and team scanning
Longxia is evaluating local/private source analysis, update monitoring, policy gates, and team workflows. No pricing has been announced.
- Private repositories and local bundles
- Commit and permission drift monitoring
- CI checks and organizational policies
FAQ
Good questions are a security feature.
Longxia is early, direct, and transparent about what agent security can—and cannot—do.
What does Longxia scan?+
The public beta scans supported text files in a public GitHub repository, folder, or file. That includes Markdown instructions, JavaScript and TypeScript, Python, shell and PowerShell, JSON, TOML, YAML, manifests, and common lockfiles.
Is this only for OpenClaw?+
No. Longxia looks for risky behavior in source and instruction files, so the same scan can help review repositories made for OpenClaw, Claude Code, Codex, Cursor, or another agent. It does not claim full runtime coverage for those platforms.
Can I scan a private repository?+
Not in the current public beta. Longxia accepts only public GitHub URLs and has no account or GitHub authorization flow. Private and local scanning are roadmap items, not available features.
Can a scan guarantee a skill is safe?+
No. Longxia is a rule-based, point-in-time static scan. It can miss malicious behavior and can produce false positives. Use the report as evidence for manual review, sandboxing, least privilege, and other security controls.
Your agent moves fast.Trust should keep up.
Scan your first agent skill free and see exactly what you are about to trust.
Public GitHub only · 10 scans per day · Reports expire in 30 days