Tools provided
Tools the agent exposes — your AI client calls them automatically when it needs them.
- open_form — Use this the moment the user wants a page audited for AI-slop UI patterns and has not pasted the source yet. It returns the questions to ask (the page source, an optional brand font list, an optional screenshot) so a form can be shown.
- run_form — Internal: invoked when the user submits the form opened by open_form. Not for direct use. Same fields and same result as run_full.
- run_full — Run the whole audit end to end for one pasted page, the same as audit_page. Use when the user pastes a page and wants the full answer, including questions like is this page accessible, WCAG compliant or legally OK: it gives measurements only, no conformance or legal verdict. Missing the page source, it asks for it instead of failing.
- audit_page — Audit ONE page's pasted HTML (and optional CSS) for nine AI-slop UI patterns: fonts, purple-blue gradients, nested and glass cards, heading skips, missing alt, small text, low contrast, filler text. Rule id, line, selector, quote and reason per finding; rules that could not run say why. No score. Use it for is this page accessible, WCAG compliant or legally OK: measurements only, no conformance or legal verdict. Screenshot notes only when the server has state signing configured.
- compare_pages — Compare two versions of a page (html_before and html_after, optional CSS for each) with the same nine rules and sort the findings into fixed, introduced and unchanged, matched by rule, selector and evidence, not by line. Use it to answer did my fix work. Missing either version, it names the missing field.
- list_capabilities — List the nine rules with their severity and what each measures, the input limits, the refusal codes and what this agent does not do. Use for discovery before an audit.
- plan_inputs — Plan the inputs for a tool before running it: returns the questions, the schema and an editable example with empty values. Use when the user is unsure what to provide.
- discover_intent — Use when the user states a goal in their own words (why does my landing page look generic). It restates the goal, asks the few questions that pin the input down and proposes an input built only from what was provided, to confirm before running.
- get_flow_status — Panel-only status probe: the result panel asks it where the last audit stands. Not for the model. It answers the last flow this server recorded, or a fresh all-pending one; hosted sandboxes are per call, so it may answer pending in hosted mode.
