A small site should leave a small data trail.

This policy applies to the public website at https://mordred.ai, including its static pages and optional hosted Mordred lab. It does not apply to websites operated by GitHub, PyPI, or another third party, even when this site links to them.

In this policy, “we” means the maintainers of the hermes-mordred project.

The hermes-mordred software sends us no usage or diagnostic data. Its audit log stays on your device unless you choose to share it; it starts as mode-0600 plaintext NDJSON and becomes MRAL-encrypted after keyvault audit-key provisioning. Any service you configure the software to contact is governed by that service's own privacy terms.

The site has no account system, checkout, comments, mailing list, or contact form. Outside the optional lab, its code does not ask for your name, email address, account details, message content, or precise location.

Like any website delivered over the internet, a visit creates a network request. Our hosting and security provider may process the IP address, requested URL, date and time, user agent, and routing or security metadata needed to deliver the page and defend the service.

Technical request data is processed only to deliver and cache pages, keep the site available, detect and prevent abuse, investigate operational errors, and meet legal obligations. We do not use it to build advertising audiences or individual browsing profiles.

When you run a policy test, the selected policy and scenario are processed inside the disposable container. The encryption proof generates and removes a random test file inside that container. If you choose to use Hermes Chat, your message and the generated response are processed to provide that conversation. Content-free counters enforce the session and daily limits.

Where data-protection law requires a legal basis, this processing supports the legitimate interests of operating a secure, reliable public documentation site. Those interests are limited by the data-minimising design described on this page.

Starting the lab creates a pseudorandom session identifier, an expiry time, and content-free usage counters. The browser receives an encrypted HttpOnly capability cookie that expires within five minutes; JavaScript cannot read the underlying session token.

The guided test uses fixed text and a controlled non-network transport. Mordred's real Strict hook refuses the request before any external model is called. Other policy probes likewise return only decision codes and content-free audit lines.

The encryption proof request contains no visitor fields. A disposable subprocess generates a random challenge, creates encrypted test files, returns content-free cryptographic checks, and removes its files and mutable key buffer. It does not call an AI model.

Hermes Chat is optional and separately labelled. It accepts at most three messages of 200 characters per session. If you use it, the message you type, recent conversation context, and the generated response pass through our Cloudflare-hosted components and remote Cloudflare Workers AI. The application keeps that conversation only in browser memory and container RAM, does not write it to a visitor account or application database, and destroys the container when ended or expired.

Cloudflare may process model input and output plus infrastructure and security data under its service configuration and privacy terms. Do not enter personal data, credentials, wallet material, or other secrets: Mordred controls destinations but does not scan or redact message content.