Teardown takes a live website URL + a plain-English question + a focus lens and returns a deep engineering architecture spec as a published HTML artifact. This page explains how it works end to end.

The estimate → run → artifact flow

Estimate. You enter a URL and focus lens; a fast heuristic estimator (no LLM) fetches the site's public surface and returns a credit range:

{ url, kind, routes, jsChunks, jsBytes, credits_min, credits_max, focus }

Nothing runs until you press "Run teardown". Estimates are never auto-charged.

Run. The engine pipelines five stages — recon → plan → dissect → synthesize → build:

• recon: fetches the public surface (HTML, JS chunks, CSS), fingerprints frameworks, extracts backend-surface hints and string literals into facts.json / strings.txt. Public assets only; authorized-recon rules, no credential probing.
• plan: ONE cheap model call turns facts + your question + lens into a 6–12 section task list tailored to the target.
• dissect: fans out one model call per section with bounded concurrency, retries, and resumability (finished sections are skipped on re-run).
• synthesize: ONE strong model pass writes an executive summary plus a "how to clone this" section.
• build: compiles sections + summary into a single self-contained, mobile-first index.html.

You can poll progress per stage while it runs (usually a few minutes).

Artifact. The finished spec is published at a stable URL and linked from your history alongside the question you asked, the credits spent, and the date.

Focus lenses

The lens changes both section selection and cost (non-everything lenses are cheaper — see pricing).

Credits

Credits are the display unit: 1 credit = $0.01. Your balance is shown in the header after sign-in. A run is gated against the estimate's upper bound; after completion you are charged for what was actually analysed (credits_spent in the run metadata). New accounts get a free 100-credit bucket. See pricing for tiers.

HTTP API

All JSON, bearer-token or httpOnly-cookie session auth:

• POST /api/teardown/estimate {url, focus} → credit estimate. No auth.
• POST /api/auth/request {email} → magic-link sign-in.
• GET /api/auth/verify?token=... → sets session.
• GET /api/me → {email, credit_balance}.
• POST /api/teardown/run {url, ask, focus} → {id} (auth).
• GET /api/teardown/:id → status/pct/artifact_url (auth).
• GET /api/teardowns → your teardown history (auth).

An OpenAPI description is served at openapi.json; agents may also prefer this page's markdown twin (/docs.md or Accept: text/markdown).

Privacy & authorization

Teardown fetches only the public surface a normal browser would receive. No credential guessing, no exploitation, no private data. Runs and history are private to your account.