Geiger

Why it's ranked here

The case for this listing is architectural discipline paired with breadth. Zero dependencies means nothing new gets installed to audit what's already installed, and every scan is read-only except the report file you name. Coverage runs wide, not just Claude Code and Cursor but a long tail of CLIs, JetBrains plugins, and AI-native browsers. Secrets are matched by regex shape and never printed, with a redaction pass applied to all output as a second layer. Releases build from GitHub Actions with signed npm provenance, so the published package traces to the exact commit it came from. Small honest surface, wide coverage: that combination is uncommon in this category.

What's good

Findings carry origin as well as exposure: a registry package, a local script, or a remote URL, each labeled by how far you can trust it. A local script tied to nothing published gets its own separate flag rather than blending into an ordinary dependency. Policy wrappers in front of an MCP server are unwrapped, not hidden: the tool reports the real server and the enforcement layer as two findings instead of one. A config file that fails to parse still gets scanned for credential shapes and flagged when it looks like it declares servers or hooks, instead of being silently skipped.

Tradeoffs

The scan only checks known locations, so an agent in a nonstandard path, inside a container, or reached through WSL from the Windows side won't show up. It reads configuration, not behavior: a finding says what a plugin's position allows, not what it did. Origin tracing stops at where a package came from; a registry-published server can still be malicious, and geiger can't tell you that. AI browsers and desktop apps are flagged by presence at medium confidence rather than parsed configuration, and the macOS paths for two of them are described as best-effort guesses. Coverage of any given tool depends on someone having added a detector for it.

How to use it well

Run it once, read the plain-language exposure tags, remove anything unrecognized, and rotate any credential it flags. Save that clean state as a baseline JSON file, then run again with diff and strict enabled from cron or CI: the job fails only when something new can execute code or hold a secret, not on the inventory you reviewed. For a fleet, the same JSON output per machine turns individual audits into a queue of ongoing drift. It tells you what is installed and what it can reach, not whether any of it is safe, so treat a clean scan as a starting inventory, not clearance.

Observed

License

MIT

Runtime dependencies

Zero declared dependencies in package.json

Language

JavaScript, published as an ES module (type: module) for Node.js 18 or newer

Interfaces

CLI only, run via npx geiger-scan; no library, MCP, or web API surface

Platform support

Windows, macOS, and Linux path tables defined in src/platform.js

Output formats

Prints a terminal report by default; writes a file only when --json or --html is passed

Detector architecture

Ten ecosystem detectors registered in src/detectors/index.js, each run inside a try/catch in src/engine.js