FAQ
Common questions about the free scanner, how to install it, and what is actually proven today.
Is Hermes Shield available today?
The free scanner is live: run `pip install hermes-shield-scanner` then `hermes-shield demo`. It runs locally on Python 3.10+ and is install-verified on Linux, macOS and Windows. It already writes a machine-readable report (`./shield-report/outputs/hermes_shield_report.json`) you can gate a CI pipeline or an AI agent on today — parse the JSON, since the exit code is always 0 for now. The paid Repairer daemon is coming.
What is proven today?
An independent reviewer confirmed ten protected Hermes live-action paths, and its discovery engine has mapped over a thousand action surfaces across the customer-zero repo — with more found as it runs.
Does it make my agent fully safe?
No tool can promise that. Current proof coverage is bounded to the tested threat model. Residual risk is documented.
Does my data leave my machine?
Hermes Shield is local-first by design: controls, approvals and the audit trail run where your agents run. It does not depend on a remote service being reachable.
How is this different from a prompt filter?
Prompt filters inspect text going into the model. Hermes Shield controls the action side: what the system is about to do — send, post, write, call — is verified at the moment of execution, so a compromised model becomes a contained failure instead of a live incident.
What agents or frameworks does it apply to?
The action-control model applies to any agent that can act — browse, email, post, approve, schedule, write files, trigger APIs. The free scanner maps that blast radius today: Python gets the full engine, TypeScript/JavaScript and C# get detection-only (via the `[multilang]` extra), and other languages via the optional `--semgrep` comparator.