Can agents find your site and its capabilities?
The page registers no tools via the WebMCP browser API (navigator.modelContext).
Fix: Expose your actions to AI agents with the WebMCP browser API: register tools via navigator.modelContext.registerTool. Optionally also publish a /.well-known/webmcp.json tool catalog as a discovery signpost.
Crawled 3 of 12 discovered pages beyond the homepage.
Lighthouse Agentic Browsing audits ran.
Could not identify a known website platform (custom stack or unrecognised).
Can agents understand your content?
Homepage ignored Accept: text/markdown and returned HTML.
Fix: Serve a markdown version of high-value pages when the client asks for text/markdown. AI summarisers, chatbots and IDE agents prefer markdown — fewer tokens, no DOM noise.
Only 8/45 images have width+height — large CLS risk.
Fix: Set width and height on every <img> so the browser reserves space before the image loads. A layout that shifts while loading makes an AI agent mis-click the element it targeted.
Homepage server response contains 1088 words of visible text — content is reachable without executing JavaScript.
Homepage contains ~1111 elements.
49 headings, no level skips.
Language declared as "en-US".
45 of 45 <img> tags have alt attributes.
101 internal links found.
All four landmark elements present (main, header, footer, nav).
Exactly one <h1> on the homepage.
All 3 reachable key pages have both a <title> and <h1>.
All 3 pages return ≥80 words of visible text without JavaScript.
Can agents actually operate the page?
Lighthouse flagged this audit — agents may be unable to perceive or operate the affected elements.
Fix: Remove ARIA attributes that aren't allowed on the element's role.
Lighthouse flagged this audit — agents may be unable to perceive or operate the affected elements.
Fix: Give every <button> discernible text or an aria-label so agents can identify it.
The page registers no tools via the navigator.modelContext browser API, so an AI agent has no structured way to operate it.
Fix: Expose your site’s actions to AI agents with the WebMCP browser API: call navigator.modelContext.registerTool({ name, description, inputSchema, execute }) from your page so an agent can invoke them. Implement it directly, or with a library like the @mcp-b polyfill (https://mcp-b.ai). Spec: https://github.com/webmachinelearning/webmcp.
2 of 10 agent-critical accessibility audits failed (Buttons do not have an accessible name, `[aria-*]` attributes do not match their roles) — an agent may be unable to identify or operate the affected elements.
Fix: Fix the failing agent-accessibility audits listed below (accessible names on controls, valid ARIA roles/relationships, nothing interactive hidden from the tree).
Lighthouse accessibility score: 83/100. Agents read the page through its accessibility tree, so this is a proxy for how navigable your site is to an AI agent.
Fix: Resolve the failing accessibility audits below — each one removes an element or relationship an agent would otherwise be blind to.
No autocomplete attributes on any data-collecting form field.
Fix: Add autocomplete hints (email, tel, name, etc.) so form-filling agents and browsers can populate fields without guessing.
0/1 data-collecting forms expose errors via ARIA attributes.
Fix: Wire inputs to their error containers with `aria-describedby` and toggle `aria-invalid` on validation failure. Agents that can't see styled error text rely on these attributes.
No HTML5 validation hints (pattern / minlength / maxlength / min / max) on any data-collecting form input.
Fix: Publish input constraints via HTML attributes so agents and browsers can validate client-side.
`[aria-hidden="true"]` is not present on the document `<body>` passed.
`[aria-hidden="true"]` elements do not contain focusable descendents passed.
`[role]`s have all required `[aria-*]` attributes passed.
`[role]` values are valid passed.
`[aria-*]` attributes are valid and not misspelled passed.
`[aria-*]` attributes have valid values passed.
Document has a `<title>` element passed.
Links have a discernible name passed.
3 of 3 visible interactive form fields have labels.
1/1 data-collecting forms mark required fields.
0.020 (field data from CrUX) — good.
651ms (field data from CrUX) — good.
No tool catalog was found at /.well-known/webmcp.json (or /.well-known/webmcp). The catalog is a community convention, not part of the WebMCP standard, so this does not affect WebMCP presence.
Fix: Consider also publishing a tool catalog at /.well-known/webmcp.json: a JSON document with a "spec" of "webmcp/0.1" and a "tools" array, where each tool declares a name and a clear description. It is a community convention (optional, not part of the WebMCP standard) that lets crawlers and agents discover your tools without executing JavaScript.
Can agents safely transact?
/.well-known/oauth-authorization-server returned 404.
Fix: Publish /.well-known/oauth-authorization-server so AI agents discovering your OAuth setup can negotiate flows automatically. Required if your site offers an authenticated API.
/.well-known/oauth-protected-resource returned 404.
Fix: Publish /.well-known/oauth-protected-resource so AI agents discovering your OAuth setup can negotiate flows automatically. Required if your site offers an authenticated API.
No Web Bot Auth signature headers — sites can't verify agent identity.
Fix: Web Bot Auth (IETF HTTP Message Signatures over Signature / Signature-Input) lets you cryptographically verify which agent is hitting you. Several CDNs offer turn-key support; otherwise skip until vendor support matures.
No Content-Security-Policy header on the homepage.
Fix: Add a Content-Security-Policy. Even a strict default-src directive cuts XSS blast radius dramatically. Start in report-only mode to find violations.
No Referrer-Policy header on the homepage.
Fix: Add `Referrer-Policy: strict-origin-when-cross-origin` so outbound links don't leak full URLs (including query strings) to third parties.
max-age=31536000, includeSubDomains=true, preload=false.
Homepage scheme is https:.
http://hetzner.com → https://www.hetzner.com/ (1 hop).
No stack-trace markers in the first 5KB.
Homepage HTML references no http:// resources.
Valid for 192 more days (Jan 29 23:59:59 2027 GMT, issuer: Thawte TLS RSA CA G1).
Server / X-Powered-By headers don't leak product version.
X-Content-Type-Options: nosniff.
X-Frame-Options: SAMEORIGIN.