Methodology

Verune finds public AI-agent tools, downloads their published code where it can, and examines it with static analysis. Nothing is ever executed.

Where tools come from

Crawlers read the MCP registry, the Claude plugin marketplace, npm, PyPI and GitHub topics on a schedule. Pro and Team members can also submit repositories, which an operator reviews before anything is downloaded. Eligible public records are listed without per-record manual review.

How coverage is measured

Every tool shows how far inspection got for its latest version:

1. Listed. Found in a public source; no code downloaded yet, or none is published (remote endpoints, for example).
2. Fetched. The published artifact was downloaded, but analysis did not run.
3. Partial. Some analyzers ran and some did not complete.
4. Complete. Every static analyzer ran on the artifact.

Low coverage is not a warning sign; it means we know less. A missing finding is never a safety certification.

What the code does

Capabilities such as running shell commands, making network requests or writing files are pattern matches in the code. We show only confident matches from the current analysis. A capability is not wrongdoing: many tools need these to do their job.

Findings and reports

Detailed reports list findings from automated static analysis with severities and explanations, plus changes observed between versions, such as published content changing without a new version number or a new network destination appearing. Findings are not human-reviewed.

Static analysis does not establish runtime behavior, cover every execution path, or assess what remote endpoints do. Change records do not prove exploitation or intent.