Manifest

Manifest reads a page through its accessibility tree, the same structure assistive tech uses, then cross-references the DOM for the details that tree misses: required flags, input types, placeholders, and disabled states. The result is a description of a page's actions that holds up better than either source alone.

The core output is a JSON actions array. Each entry carries an id, label, type, description, and a requires field. That field is a set of groups joined by AND, where each inner group is a set of alternatives joined by OR. So a size picker shows up as one group with several ids, and a login button shows up as two one-item groups for email and password.

Every action ships with a locator: a CSS selector, an ARIA role, and an accessible name. The docs suggest leaning on role and name rather than CSS wherever you can, since those survive redesigns that would wreck a hardcoded selector.

You can reach Manifest two ways. Add the io.omfang/manifest MCP server to your client config and it exposes a streamable-http endpoint, or call the REST API directly with a key in the x-api-key header. Same response shape either way.

Cached responses come back fast and count against your limit less than cold requests do. Since agents tend to hit the same handful of URLs over and over, this keeps cost and latency predictable across a run.

An optional country field sets the scan's Accept-Language header and, on some sites, a locale cookie. It's best-effort: Manifest won't override geo redirects universally or convert prices, and it hands back a locale_mismatch flag so you know when the page you got was locale-specific.