Tavus screen share: the agent sees your screen; Sume takes images

Tavus screen share lets a live agent look at what a user shares and answer questions about it, but only with perception_model set to raven-1, inside a live video conversation, after a participant starts the share. Sume has no live viewing. An avatar video takes images you supply in advance: an optional product_image, a scene photo, or a per-scene background image.

So the choice is whether the agent needs to see something that changes during the conversation or you know the visual in advance.

What Tavus requires

Per Tavus, screen perception requires Raven, a live video room (not chat-only or audio-only) and a participant who starts sharing through the browser prompt. Raven can read text, spot errors and answer prompted visual questions about the shared screen. For structured goals, you use objectives with a visual modality and visual_awareness_queries. Screen share is enabled by default, and you restrict it by leaving the share button out of your UI (Tavus docs: Screen Share).

The Sume inputs

Sume's avatar video accepts product_image for a product or reference image, scene with a prompt or a photo for scene direction, and per-scene image backgrounds in video_inputs. All media fields must be fetchable public HTTPS URLs; localhost, private-network, non-HTTPS and signed or private URLs are not accepted (Media inputs).

For a software demo, export the screen states you need as stills, host them at public URLs and assign one per scene, or record the demo separately and join it to the avatar clip with Timeline.

• Check a finished clip with video_inspect: it returns stills from the video so you can confirm the right image appears in the right scene.
• Do not put a private dashboard in an image you publish; the image URLs must be public, and the finished clip can be shared by URL.

Practical preparation

For a rendered clip, prepare images at the aspect ratio of the final video, host them at stable public HTTPS URLs, and keep the originals. A product shot on a clean background works as a product_image; a lifestyle photo works as a scene.

For a live session, test the share prompt in each browser your users have, and decide in advance what the agent should say if the user declines to share.

What can go wrong

Sume's media inputs docs say signed or private URLs are not accepted, so test each URL from outside your network before you submit. A very busy screenshot may read poorly at phone size, so crop to the area that matters.

A live share can expose more than you meant to show, because the user chooses what to share. A prepared image cannot, which is one reason regulated teams prefer scripted clips.