Make a video from a URL in Claude or Cursor: crawl_scrape, Wan 3.0

To make a video from a URL inside Claude or Cursor, connect Sume's hosted MCP server, ask the agent to read the page with crawl_scrape, then to call generate_video with wan-3.0 and a prompt built from what it read. Alibaba's Wan 3.0 README (read 2026-10-05) says Wan 3.0 accepts webpages as reference assets; Sume's catalog entry for wan-3.0 says web_url is not exposed in v1. The agent does the page reading that the model would do on Alibaba's side, and hands the model text and images instead.

Connect

The MCP quickstart gives the production URL, https://mcp.sume.com/mcp, and for Claude Code the two commands below. Use OAuth, not keys pasted into chat. Cursor takes the same URL as a remote server entry.

Then ask the agent to call tools_list or mcp_health to confirm the connection.

claude mcp add --transport http sume https://mcp.sume.com/mcp
claude mcp login sume

The calls in order

The tools and gates page is the source for which tools are free and which spend money. Crawl tools such as crawl_scrape are in its Read list. generate_video is a paid tool: it needs an idempotency_key, and if you omit payload.model it routes to sume/auto unless you named a model family. So say "use wan-3.0" in the instruction, or the agent may let Auto choose.

An instruction that works

Ask for the plan before the spend. The docs suggest asking the agent to call tools_schema and explain idempotency_key and dry_run before submitting any paid generation. A request in that spirit:

Read https://example.com/blog/our-launch with crawl_scrape.
Summarize it into three visual beats, 8 seconds each.
Call tools_schema for generate_video and show me the payload you
plan to send, with payload.model set to wan-3.0, duration 24,
aspect_ratio 9:16. Wait for my OK before submitting.

Limits

The hosted MCP server cannot read files on your laptop. If you want your own photos as references, the docs describe an upload flow: create an upload URL, PUT the bytes from your client, then complete the asset. For URLs of images already on the page, the agent can pass them straight in as reference images, up to the 10 allowed by the catalog entry. Whatever the page says, the model only sees the prompt and the references you send.

Keep one habit from the docs: use tools_list and tools_schema for the live contract, because the docs warn not to assume parity with the HTTP API. If a field the agent proposes is not in the schema it returns, drop it. And if the job takes longer than one wait, ask the agent to wait again instead of resubmitting; a second submit with a new idempotency_key is a second charge.

For a long page, ask the agent to keep the prompt to the three or four ideas that can be shown in pictures. A 2,000 word blog post will not become a 2,000 word prompt, and the clip will be better for the cut. Ask the agent to list what it left out so you can decide whether a second clip is warranted, and to name the image URLs it chose as references so you can swap any that show logos or people you do not have rights to use.