Agent discovery

CoreX publishes a few machine-readable files for the surfaces it actually serves. A file is not published for a protocol CoreX does not implement.

What is public

URL What it is
/robots.txt Crawl rules. Content-Signal sits in the User-agent: * group.
/llms.txt Short index of the public site, the docs, the free tools, the REST API, and the MCP server.
/llms-full.txt That index, then the text of the key public pages and the live docs. Both files are cached.
/sitemap.xml Sitemap index.
/openapi.json OpenAPI 3.1 for GET and POST /api/v1.
/.well-known/api-catalog RFC 9727 linkset (application/linkset+json) pointing at GET /api/v1 (a short JSON index), /openapi.json, and this API's docs.
/.well-known/mcp.json MCP server card for POST /mcp. The same card is at /.well-known/mcp/server-card.json.
/auth.md How an agent sends an API key, and the actions the REST API and MCP server expose.
/agents.json The same note as JSON, generated from the OpenAPI document and the MCP tool list.

There is no public blog. /blog is not a page.

A public marketing, docs, free-tool, or leaderboard URL accepts Accept: text/markdown and also answers at <url>.md (the home page is /index.md). The response is text/markdown; charset=utf-8 with Vary: Accept. The header does not redirect and does not apply to /app. Docs already had /docs/{page}.md.

Access-Control-Allow-Origin: * is on those discovery files, on GET /api/v1 (the index only), and on markdown responses. It is not sent with credentials, and it is not added to /api/v1/* (the keyed REST API), to a 404 under /.well-known/, or to session and panel routes. The widget, ingestion, and embed endpoints keep their own wildcard because a merchant page calls them, except a path under those prefixes that matches no route: that answer is JSON not_found and carries no CORS header. /openapi.json is a file in public/, so the production web server sets Content-Type: application/vnd.oai.openapi+json, the CORS header, and a 204 for OPTIONS before PHP sees the request. The same layer adds Strict-Transport-Security, X-Content-Type-Options, X-Frame-Options and Referrer-Policy only when the path is a real file, and it does not gzip woff2, PNG, JPEG, GIF, WebP, AVIF, ICO or a file that is already compressed. PHP responses stay as the app set them, so a booking or feedback page can still be framed. GET /free-tools/leaderboard is an edge-cached read.

Markdown for pricing, a category leaderboard, and State of the App Store includes the page's facts (what the beta includes, the top of the ranking, the month's launch figures), not only the title and a link list. The HTML render still records Vite chunks, and the preload middleware writes them into a Link header before the body is replaced. The markdown middleware is the outermost one, so it drops rel=preload and rel=modulepreload after that write. The same drop covers /llms.txt and /llms-full.txt. A cache that stored the header would answer the next load with 103 Early Hints.

Docs are English. /fr/docs keeps lang="en", does not advertise a French hreflang, stays noindex, and sets rel="canonical" to the English document. French pages link to /docs. On a French page that does have a translation, JSON-LD url, @id and breadcrumb item use the same /fr URL as the canonical. /privacy.md and /terms.md are the legal pages' own copy, and a docs Accept: text/markdown response is the same bytes as /docs/{page}.md. OPTIONS /api/v1 and OPTIONS /api/v1/* answer 204 when the path exists and the not_found envelope when it does not. robots.txt also disallows /fr/free-tools/*/run.

What is not public

The API and the MCP server use an API key (Authorization: Bearer or X-CoreX-Token). CoreX does not run an OAuth authorization server, so it does not publish protected-resource or authorization-server metadata.

There is no Agent-to-Agent endpoint, so there is no /.well-known/agent-card.json. poppy.json is not a protocol this app speaks. DNS-AID is a DNS record; nothing in the application can publish it.

Organization.sameAs is only the URLs in ORGANIZATION_SAME_AS. The default list is empty, and an empty list omits the key. Do not add a profile URL that the product does not already link.

Content-Security-Policy

Responses send Content-Security-Policy-Report-Only unless CSP_REPORT_ONLY=false, which switches the same policy to the enforcing header. CSP_ENABLED=false removes it. X-Content-Type-Options: nosniff and Referrer-Policy: strict-origin-when-cross-origin are always set.

The policy allows the Vite build on this origin, inline JSON-LD and Inertia data, PostHog, Cloudflare Web Analytics, Mailchimp, Google Analytics hosts (nothing in the app injects a tag today), Shopify CDN images, the self-hosted fonts, and the same-origin live-chat widget and embeds. CSP_EXTRA_SCRIPT_SRC and the other CSP_EXTRA_* variables append hosts. The report-only policy always names POST /csp-report (report-uri and report-to). That endpoint rate-limits, logs the directive and the URL path, and drops query strings. CSP_REPORT_URI (or SENTRY_CSP_REPORT_URI) adds a second https collector. Leave the policy report-only until those reports show it is not blocking a real asset.

Deploy and rollback

No migration. llms.txt and llms-full.txt use new cache keys, so a deploy does not need a cache flush. Rollback is a revert of this release. If an enforcing content-security policy blanks a page, set CSP_REPORT_ONLY=true (the default) and redeploy the config. Clearing ROBOTS_CONTENT_SIGNAL removes the Content-Signal line. Clearing ORGANIZATION_SAME_AS removes sameAs.