- Home
- Skills
- Agents & Orchestration
- skill-manager
Skill Manager
Secure manager for skills featuring whitelisted GitHub/npm installation and AST-based security auditing.
Free
Works with the AI Agents you already use
skill-manager
Example session with this skill installed
Install the meeting-notes skill from GitHub into my local ~/.dsh/skills library, run a static security audit on it, and report the audit level with any high/critical findings before I use it. Show the install source, the audit result, and the recommended next checks.
- Read your context and instructions
- Compiled the -manager
Summary
Installed the meeting-notes skill from a whitelisted GitHub source into ~/.dsh/skills/meeting-notes and ran a read-only static audit. Sample result: level safe, count: 0 findings. This is a heuristic rule-based check: it means no rule fired in this sample, not a certification that the skill has no vulnerabilities.
Installation Record
| Step | Detail | Status |
|---|---|---|
| Source | github:owner/meeting-notes; host allowlist (codeload.github.com) | Accepted |
| Network | HTTP 3xx redirects not followed; non-allowlisted/internal hosts blocked | OK |
| Extraction | Entry-count and size budgets; tar-slip path-boundary validation | OK |
| Placement | Extracted to ~/.dsh/skills/meeting-notes; existing directory not overwritten | OK |
Audit Output
Command: python scripts/dsh_tools.py audit meeting-notes --json
{
"name": "meeting-notes",
"level": "safe",
"count": 0,
"fail": false,
"fail_on": "",
"findings": []
}
Connects securely to your tools. The creator never sees your data.
About this skill
Managing third-party AI agent skills is a security minefield of prompt injections, credential theft, and unauthorized script execution. This package provides a secure infrastructure to discover, install, and audit DeepSeek Harness (DSH) compatible skills. It enforces strict domain whitelisting for downloads and implements 12 distinct security audit rules, including AST-based Python analysis, to prevent data exfiltration and shell-pipe attacks before a single line of text is loaded into your agent's context.
What it does
- Secure installation downloads SKILL.md packages from whitelisted GitHub and npm sources with SSRF protection.
- Static security auditing detects prompt injection, credential phishing, and dangerous shell commands using AST and pattern matching.
- Skill discovery lists all locally installed capabilities with their specific triggers and metadata.
- Plugin identification distinguishes between pure Markdown-based skills and TypeScript-based DSH function plugins.
- Safe loading provides a read-only preview of skill instructions without executing them, preventing unauthorized invocations.
How it works
- Locate existing skills or identify new local directories using the
listandinfocommands. - Install remote skills via
github:owner/repoornpm:packageidentifiers with built-in tar-bomb protection. - Audit the package using
auditto check for high-risk findings likerm -rfor network exfiltration patterns. - Load the audited content into the agent context for task execution.
Frameworks & tools
Built for the DeepSeek Harness (DSH) ecosystem. Requires Python 3.9 or higher (standard library only). Works with any LLM agent capable of reading Markdown files and frontmatter.
Why this beats prompting it yourself
Manual auditing of complex agent instructions is prone to human error and ignores hidden malicious scripts. This tool automates the detection of exfiltration patterns and dangerous shell commands that simple system prompts cannot reliably catch.
Use cases
- Auditing a third-party GitHub repository for hidden prompt injections or credential-stealing instructions.
- Centralizing a library of custom agent skills in a standard local directory (~/.dsh/skills).
- Safely downloading and extracting SKILL.md files from npm without manual zip handling.
- Verifying if a local directory is a function-based plugin or a text-based skill.
Known limitations
Static auditing is heuristic and may have false positives or negatives. Binary files and files over 1MB are skipped during security scans.
Auditing is heuristic and may miss sophisticated obfuscated attacks. Skips binary files and files >1MB. Requires Python 3.9+ for full AST-based security checks.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before any payment.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
Fresh listing
Recently published to Agensi
Free forever
No account required to browse
Trust & safety
Security scan
Security scan passed 6 days ago
- Free with an Agensi account
Requires
Permissions
Allowed Hosts