Start with the distinction
This website is a self-contained product preview. The planned operational platform is separate. The browser demos do not collect security telemetry or execute defensive actions.
The proposed control loop
The intended sequence is: authorised source → validated event → source-aware entity and relationships → detection → incident → evidence-backed explanation → scoped policy → approval → executor → verified audit outcome.
The AI tool boundary
An enforceable gateway must verify an agent’s identity and the exact downstream resource. Restrict bypass routes and direct credentials. Prompts or SDK wrappers alone cannot create a reliable boundary around an agent with unrestricted access.
Guardian modes
Observe is read-only. Assist proposes a specifically scoped action for approval. Autopilot is intended for narrow, validated, pre-approved actions only. Automated response is not a licence for the model to change organisational policy.
Using this preview
Try the console tabs on the homepage, select graph nodes, run the local threat scenario, change AI policy requests and play the incident replay. The command menu opens with Control K or Command K. Reduced-motion settings are respected.
APIs and installation
No production API contract, SDK package or endpoint installer is published in this release. There is no real “npm install omniguard” instruction. Any illustrative interface must not be treated as a supported production integration.
Hosting this website
The supplied prebuilt files are designed for Cloudflare Pages Direct Upload. The package includes a small optional Cloudflare Worker for early-access submissions and a deployment guide. The static website does not require a Node build, a site builder account or an AI API key.