Make AI UGC videos with Claude: avatar, voice, lip sync

To make AI UGC videos with Claude, give Claude a video tool. Claude's models take text and images in and put text out, so the clip itself comes from a tool that Claude calls. Connect Sume's hosted MCP server and allow Write at sign-in, and Claude can pick or create an avatar, voice the script, lip-sync the avatar to it, and hand back a media.sume.com link.

Claude facts come from Anthropic's models overview and custom connector article; the Sume side comes from the MCP overview, MCP tools and gates, and Jobs and results, read on 2026-09-27. Where a step comes from the tool descriptions in Sume's current code, it says so. For other kinds of video, see can Claude make videos?.

How do I connect Claude to Sume?

Add https://mcp.sume.com/mcp to Claude as a custom connector, Claude's way of connecting to remote MCP servers, and turn Write on at Sume's consent screen, because paid tools such as avatars_create need it. Sume has no official connector for Claude; Add Sume to Claude as a custom connector walks through the screens and scopes.

Sume's docs say hosted MCP still works but is not part of the primary path: reach for it when your environment calls for it, not as the default integration (Sume basics). A person working in Claude is that case; an app that makes videos for its users would call the Format API or the Developer API.

What does Claude do to make a UGC video?

Sume's docs say every on-camera speaking shot, short UGC and presenter ads included, is VEED Fabric 1.0 with an accepted still plus text-to-speech. The avatar tool descriptions in current code steer Claude through the same steps:

How do I ask Claude for the video?

Describe the ad the way you would to an editor, and name the guardrails. Claude writes the script itself; UGC ad script for AI avatar videos covers the hook, demo, and call-to-action structure. For example:

Make a 20-second UGC-style video for our steel water bottle with the Sume tools.
Find a friendly presenter with avatars_search, then make a still of them for this
shot with generate_image. Write a casual script: a one-line hook, one line on the
product, a call to action. Voice it with tts_create, then lip-sync the still with
avatar-image-to-video_create. Before each paid tool, run it with dry_run=true and
show me the estimate. Set max_spend_usd on every paid call. Wait with jobs_wait,
then give me the link.

How do I keep Claude from overspending?

Paid tools spend from your Sume wallet, subject to Sume's admission checks, at API pricing rates: text-to-speech is $0.0475 per 1,000 characters, and VEED Fabric 1.0 is $0.1875 per audio second (720p), plus a 5.5% agent fee by default. In current code, the descriptions of avatars_create and avatar-videos_create also tell the agent to "Confirm with the user before every paid create." The hosted tools add three gates:

• idempotency_key is required on write and paid tools: a stable key that deduplicates a repeated call.
• dry_run=true returns the admission and cost preview without submitting.
• max_spend_usd caps a call's spend, and is enforced only when provided.

What are the limits?

• Hosted MCP cannot read files from your laptop. Give Claude public HTTPS links for photos, such as a product shot or the face for a new avatar.
• Fabric takes up to 300 seconds of audio. The audio must sit on Sume's media host, as tts_create output does, and be at most 10 MB.
• One jobs_wait call holds at most 55 seconds. The docs say to repeat the wait on the same job, never to resubmit the paid create; MCP timeouts on long video jobs explains why.
• The direct Avatar 1.0 talking-video tool, avatar-videos_create, is marked in current code as a legacy path for explicit requests, and its scripts must be English.