Canonical: https://botbento.com/usecases/vet-third-party-ai-tool-packages-before/
Format: Markdown representation of the public HTML page.

# Vet third-party AI tool packages before installation

[Operations](https://botbento.com/usecases/ops/) · Intermediate

Compare declared capabilities with actual files and probe risky servers.

A bot reviews an AI tool package or server repository before adoption. It maps declared features and permissions, inspects files and configuration for mismatches, runs the server in a controlled probe where possible, and produces a plain-language risk report. Nothing is installed or connected until the user approves.

## Quick answers

**What does it do?** Compare declared capabilities with actual files and probe risky servers. Result: Risk report with approval gate.

**What does it need?** Tools: GitHub, Terminal, Docs. Intermediate: some setup. The prompt includes an approval step.

**How do I run it?** When something happens: New tool package submitted. Copy the prompt below into a bot in BotBento.

## How it works

1. When something happens: New tool package submitted
2. Then: Inspect files and probe server
3. Result: Risk report with approval gate

## Prompt to start with

```text
When a new package is submitted, read its declared capabilities and permissions. Compare them with the repository files and configuration. Run the server only in an isolated probe and record its observed behavior. Flag mismatches, unclear claims, and risky access. Write a short risk report with evidence and recommended next steps. Ask for approval before installing, connecting to tools, sending messages, making payments, or deleting anything.
```

## Tools it uses

GitHub, Terminal, Docs

## Works with

BotBento

## Tags

security, review, installation, automation, risk

## Source

[USB: A trust layer for AI skills and MCP servers](https://news.ycombinator.com/item?id=50432661) — Hacker News · checked 2026-11-13

Summarized in our own words. Any results described there are the authors’ own, not measured by BotBento.

Verified 2026-11-13: passed every automated gate and a second-model review (kimi-code/k3.1-pro). Automated, not human review. [How verification works](https://botbento.com/usecases/verification/)

By BotBento Editorial · Added 2026-11-13 · Updated 2026-11-13 · [Suggest a correction](https://botbento.com/contact/)

## Related use cases

- [Fraud rule drift review and variant selection](https://botbento.com/usecases/fraud-rule-drift-review-and-variant-selection/) — Check rule performance, spot drift and propose the best variant
- [Pre-flight review of autonomous coding agent shell commands](https://botbento.com/usecases/pre-flight-review-of-autonomous-coding-agent/) — Analyze proposed terminal commands to flag destructive operations before execution
- [Scan repositories for committed local settings secrets](https://botbento.com/usecases/scan-repositories-for-committed-local-settings-secrets/) — Find credential-bearing config files in source control before they spread
