How it works

Hierophant is a static checker. It reads files as text and applies a fixed, versioned set of rules. It does not use a language model, and it never runs anything from the repository.

1. Resolve the commit

Hierophant asks the GitHub API for the repository and resolves the default branch, or the branch in your URL, to one commit. Private repositories are refused.

2. Read the agent files

It downloads only files that agents load or run, such as hook settings, instruction files, MCP configs, editor tasks, dev containers, and the root package.json. It also reads local scripts that those files launch. The repository is not cloned, installed, or built. Symbolic links, submodules, and binary files are skipped.

3. Apply the rules

Ruleset v0.3.0 looks for automatic commands, remote code, concealed payloads, risky MCP servers, weakened approvals, and text that tries to steer an agent.

4. Map the impact on each agent

Each finding is matched against what the vendor documents about Claude Code, Cursor, VS Code, Antigravity CLI, and Gemini CLI. Those sources were last checked on 2026-11-22. When a vendor does not document a behavior, the report says Behavior unknown instead of guessing.

5. Check the packages

If an agent file would download a package with a tool like npx or uvx, Hierophant looks up its name and version in OSV, npm, and PyPI. It flags packages listed as malicious, packages missing from the registry, known vulnerabilities, and packages created in the last 30 days. Only package names and versions are sent. Results are renewed when a report older than a day is scanned again.

6. Save the report

A report stores metadata: the repository, commit, ruleset version, status, rule IDs, file paths, line numbers, and package results. File contents and excerpts are not stored. If the same commit was already checked with the same ruleset, you get the existing report.

Scan limits

  • Up to 200 agent files per scan.
  • Up to 256 KB per file.
  • Up to 5 MB of text in total.
  • Up to 25 findings shown per file, most severe first.
  • Up to 20 packages per scan.
  • About 20 seconds per scan.

The hosted site currently reads about 30 files per scan, starting with hooks and configurations. Anything left out because of a limit is listed in the report.