AGENT KBASE
Wizard

Start · 10 min read

Extract from website

Prefill the wizard or CLI from a public marketing or docs page.

Wizard

  1. Open /wizard.
  2. Paste https://… and click Extract from website.
  3. Review every field. Product copy (price, SLA) is not the whole company story until you confirm it.

CLI

bash
pnpm brain extract-url https://example.com --json
pnpm brain extract-url https://example.com --init ./my-app

How analysis works

  1. Fetch public HTML (size and timeout limits apply).
  2. Parse title, meta/OG, headings, and substantive paragraphs.
  3. Synthesize a company oriented description from the hero section (not meta alone when meta is a short SKU line).
  4. Optional: if OPENAI_API_KEY is set on the server or CLI environment, an LLM pass refines the JSON (method: "openai" in output). Model defaults to gpt-4o-mini; override with OPENAI_EXTRACT_MODEL.

Response includes method, warnings, and raw for transparency.

Limits

  • Public http/https only (no localhost or private IPs).
  • HTML up to 512KB, 12 second timeout.
  • Heuristic synthesis cannot invent customers or metrics not on the page.

Always keep human review before export.

What extract does *not* do

  • No Cursor skills — only brain/facts, stories, and open questions. Import skills from SkillTree or write SKILL.md yourself.
  • No RAG index — it reads one public URL, not your whole doc corpus.
  • Heuristic vs LLM — without OPENAI_API_KEY, synthesis uses headings and body text only; set the key on Vercel or locally for a second pass.

Richer results

  1. Paste the marketing home URL (not only /pricing unless that is all you have).
  2. After extract, add repo specific facts: stack, deploy, ICP, metrics.
  3. Copy founder notes from agent.md or internal docs into brain/facts/.
  4. For Auditamia-style sites, confirm sample vs live deliverable in open questions.