Reference
Install check
The install check reads a skill or an MCP server before you install it and tells you, in plain words, what could hurt you.
#Why it exists
Skills and MCP servers run with your permissions. A skill can be a few files of instructions and scripts that your AI agent follows, so a copied one can download and run code, read your keys or quietly change your agent's settings. The install check reads those files the way a careful reviewer would, and points at the lines that matter.
#What you can paste
| Input | Example |
|---|---|
| A GitHub repository, folder or file link | https://github.com/owner/repo/tree/main/skills/name |
| A SKILL.md file link | https://github.com/owner/repo/blob/main/SKILL.md |
| An npm package page | https://www.npmjs.com/package/name |
| An npx, bunx or pnpm dlx command | npx -y @scope/name |
| A raw GitHub file or any HTTPS file link | https://raw.githubusercontent.com/owner/repo/main/file.md |
The repository or package has to be public. Open it at /vet.
#What it looks at
Scripts and code, instruction files (.md, .mdc, .txt) and configuration files. It skips dependency folders and typical documentation, asset and site folders, and treats files in test and example folders as lower priority. Comments and sentences that say what not to do ("never run curl | sh") are not counted as findings.
| Limit | Value |
|---|---|
| Download size | 15 MB (link the skill folder instead of a large repository) |
| Unpacked size | 40 MB |
| Largest file read | 300 KB |
| Files read | Up to 600 |
#What it flags
Findings are grouped by what they could do to you, with the file, the line and an excerpt for each:
- downloads a script and runs it straight away, or hides code and then runs it;
- opens a remote shell, or mentions crypto mining;
- sends data to a throwaway collection service, or posts to a hardcoded Discord or Telegram channel;
- touches your private keys or saved logins, or reads folders that hold secrets;
- turns off your agent's permission prompts, or edits your AI tool's settings;
- sets itself up to run again later, or runs a script automatically when installed;
- deletes your home folder or whole disk;
- tries to override your agent's instructions, tells the AI to hide things from you, acts without telling you, or gives orders about other tools;
- hidden text: invisible characters, text-direction tricks and long encoded blobs in the instructions.
The report also lists what the skill can do (for example, run shell commands or reach the network), the hosts it talks to and the tools it asks permission for.
#Reading the result
| Result | Meaning |
|---|---|
| fail | At least one serious red flag. Read it before you install. |
| warn | Nothing serious, but items worth a look. |
| pass | No red flags in the files that were read. |