Home · Platform · Platform · The agent
An agent that cannot publish an invalid page.
Ask Noice.work for a page and it does not generate HTML, or React, or a template. It composes structured component data, validated against a compiled schema, from a closed catalogue of audited components. The validator is not a linter that warns — it is the write path. Output that fails it does not exist as a draft, let alone a page.
Structure is the safety, review is the judgement
Every AI content tool now ships a review step, and review steps matter — editorial judgement is human work. But a review step catches what a reviewer happens to notice, and a marketing team reviewing forty AI drafts a week notices less every week. Noice.work splits the two jobs: the schema system guarantees the page is structurally valid, accessible-by-construction and on-brand, so the human review is about the one thing humans are irreplaceable for — whether it says the right thing.
Composition is grounded, not improvised. The agent retrieves from your governed content library before it writes, and the sources it drew on can be attached to the document and travel with it — so “where did this claim come from” has an answer months later, which no audit log can give you.
Editing is governed the same way
Edits go through the same door. Ask for “make the second section about pricing and tighten the intro” and the agent proposes precise operations against the page's structure, dry-run validated for every audience variant before anything is written, reported back as a readable change list. Approvals, review chains and role permissions apply to the agent exactly as they do to a person, and every action lands in the same append-only audit log.
Bring your own agent
The agent surface is open. A first-party MCP server on your site's own hostname exposes 179 governed tools — search, compose, edit, publish, media, forms, fragments, experiments, redirects, brand tokens, video scripts and more — to any MCP-speaking agent: Claude, or your own. Keys are per-organisation, role-scoped and revocable; an external agent holds exactly the role its key carries, and every tool re-runs the same permission checks, schema validation, review gates and audit logging as the human UI. Handing an agent the keys is a governed act, not a leap.
Questions people ask
Can the agent publish on its own?
Only if your site's configuration allows it, and through the same gate as everyone else. On a site with a review chain, agent-composed pages queue for the same approvals a person's drafts do. Scheduled and unattended publishes are frozen snapshots, audited like everything else.
What model does it use?
Model choice is a platform configuration, not your problem to manage — and because the model only ever emits validated component data, the governance story does not change when the model does.