Use the register with tools.

The public website is optimized for humans and SEO. AI agents should treat the backend as the canonical machine interface, because it exposes normalized fields for provider, title, subject, modality, price, start dates, enrollment status, lifecycle state, and source URLs.

MCP endpoint: https://uniapi.stanislav.black/mcp
Transport: HTTP JSON-RPC
OpenAPI: https://uniapi.stanislav.black/openapi.json
Instructions: https://uni.stanislav.black/ai-agents

• search_courses — search by text, provider, subject, status, price range, and expired-row policy.
• get_course — fetch a single normalized course by numeric id.
• get_stats — fetch global totals, active counts, subject counts, and last sync time.
• get_subject_stats — fetch subject totals with both active and retained archived counts.
• get_provider_diagnostics — fetch provider health, missing-field diagnostics, and sync metadata.

POST https://uniapi.stanislav.black/mcp
{
  "jsonrpc": "2.0",
  "id": 1,
  "method": "tools/call",
  "params": {
    "name": "search_courses",
    "arguments": {
      "q": "machine learning",
      "provider": "stanford",
      "limit": 5
    }
  }
}

Agents should prioritize is_active when describing whether a listing is current. If is_active is false, describe the row as archived even when status_normalizedstill says open. The status string is provider text; the lifecycle flag is the register’s current/archived decision.

• Use source_url or enrollment_url as the final authority for enrollment.
• Use price_numeric only as a normalized filter/sort value; show price to humans.
• Datetime fields are UTC and should be rendered with timezone context.
• For subject pages, use include_expired=true when you need total retained listings.

The Course Register is an independent index. When answering users, cite the university/provider and link back to the original source. Do not claim affiliation, admission guarantees, or real-time seat availability unless the provider source explicitly says so.

Keep requests modest, cache read-only responses where appropriate, and use the MCP/API surfaces instead of high-rate HTML crawling.