Agent OS

The problem. Brand management runs on judgement calls that nobody records: whether a post was disclosed, whether a claim is defensible, and who signed off. The evidence for those decisions usually exists only in somebody's inbox.

What it does. Coordinates agents that read public or permissioned content, grade the evidence behind each claim, and surface disclosure and claims risk for human review, with a sign-off gate before anything is treated as settled. The Disclosure and Claims agent is hosted and answers directly: given a post, it returns whether it reads as paid, whether disclosure is present, and each claim it found with a category and a risk level.

Built on. A nine-table evidence model on Postgres with pgvector, an orchestrator that routes to a hosted model, a local one, or a deterministic fallback so it never hard-fails offline, and five agents. Seven containers in total, of which one — the part needing no database, no key and no local model — is deployed publicly.

What it does not establish. Public and permissioned data only, and not legal advice: it surfaces risk for a person and their counsel to read. The hosted endpoint is not a hosted engine — evidence storage, retrieval, the local model, workflows and dashboards all still require the self-hosted stack.

Public link held. The hosted address answers a single data endpoint rather than a page, so it is not offered as a link. The full source, tests and design are open in the repository.

This record is generated from the same inventory as the complete build index: the production URL was checked at the last regeneration, and the status line above repeats what the build reports about itself rather than restating an intention.