Agent turns a product page into a clip: crawl_scrape, generate_video

To have an agent turn a product page into a short vertical clip on Sume MCP, call crawl_scrape to read the page, write the brief yourself, then call generate_video. The first is a read tool, so it works on an OAuth session with Write off. The second is a paid tool, so it needs Write or an API key, plus an idempotency_key. Keeping the two in separate steps lets you review the brief before anything is spent.

The sequence

Every tool name below is in Sume's tools-and-gates inventory (read 2026-10-08). The gates column shows what the session needs.

Write the brief from the page

The scrape returns page content; the agent's job is to compress it into a prompt of two or three sentences about one product, one setting and one motion. Ask it to show you the prompt, the aspect ratio (9:16 for vertical), the length, and the dry_run estimate before it submits. With model left blank the route is sume/auto (3 to 10 seconds, 720p and 8 seconds by default), and an 8-second default clip is $1.00.

Do not paste signed URLs or tokens into the brief. If the page is behind a login, the crawl will not see the logged-in content, and the agent should say so rather than guess.

Keep spend out of the reading step

Run the reading in a session with Write off. A model that never sees generate_video cannot call it, and a stale prompt that tries gets insufficient_scope. When the brief is approved, reconnect with Write, or switch to a key-based entry, and run the single paid call with a cap.

• Use one idempotency_key per clip, such as the product handle plus the draft number.
• Use max_spend_usd just above the dry-run estimate.
• Wait with jobs_wait and repeat on wait_slice_expired; never resubmit.

Limits to know

Hosted MCP cannot read files from your laptop. If the clip should start from a product photo, the page's image URL must be a public HTTPS URL, or you upload through the documented flow: create an upload URL, PUT the bytes from the client, then assets_complete. And Sume Video 1.0 is not on MCP, so this flow uses generate_video and the router catalog.

Review checklist before the paid call

Give the agent a short checklist and have it report each item in its reply: the source URL it scraped, the one-sentence claim the clip will make, the aspect ratio and length, the dry_run estimate in dollars, the balance, and the idempotency_key it will use. A human approves in one glance, and a retry after a timeout reuses the same key instead of paying twice.