The ARV record - AgentTV

ARV (Agent-Ready Video) is an open standard: one record per video that tells AI what is in the video, who owns it and how it may play. The spec lives at agentreadyvideo.org. AgentTV builds an ARV record for every video it indexes. ARV does not replace schema.org. Search engines keep reading VideoObject JSON-LD at crawl time. The ARV record carries what JSON-LD has no place for, such as timed moments, on-screen text and rights, for agents at inference time. See Crawl time and inference time.

One record, several views

There is one record per video. AgentTV serves views of it in the places each reader looks:

View	Where	Who reads it
Page block, `head_html`	Your page head: JSON-LD plus link tags	Search engines and AI crawlers
Page block, `body_html`	Your page body: chapter summaries and transcript	People, crawlers that read visible text
ARV record	`GET /public/creators/{slug}/videos/{video_id}/arv.json`	Agents that follow the link tag
MCP tools	`https://mcp.agentcdn.com/c/{slug}/mcp`	AI apps with the connector turned on

Your page template makes one request for the page block. You do not assemble the record yourself.

What the record holds

Read a live one:

curl -s https://api.agentcdn.com/public/creators/sandersaar/videos/video_t82C_EYja18/arv.json

The record sits under data:

Field	What it is
`asset`	The video id, duration, spoken and subtitle languages (Languages), where it came from and how it plays at its origin.
`display`	Title, description, thumbnail, publish date and watch link.
`chapters`	Titled sections with start and end in milliseconds, a summary, the words said and a description of what is shown.
`moments`	Addressable spans, each with a `moment_uri` such as `arv:video_t82C_EYja18#t=0,38071` and a link that opens at the start.
`transcript_cues`	Timed sentences with speaker labels and names when known.
`visual`	Time windows with a description of what is on screen and the on-screen text.
`speakers`, `people_spans`	Who speaks and who appears, and when.
`entities`, `products`, `places`, `topics`, `labels`	What the video covers, with times. Entities link to Wikidata where they match.
`rights`	The rights in force for the video: search, cite, summarize, quote, play, clip, train and more. See Rights and access.
`model_provenance`	Which models made each layer, with versions and dates.

All times in the record are milliseconds from the start of the video. The page block’s JSON-LD uses seconds, as schema.org does.

Freshness

AgentTV rebuilds a record when a model improves or when you change a video’s details or rights. Your page gets the newest record on its next fetch of the page block. Subscribe to the video.updated webhook to know when to fetch again.

What the record never holds

A stream URL. Search and discovery never return one. A stream URL comes only from a playback call that checked the rights first. See Play a moment in an AI app.

​One record, several views

​What the record holds

​Freshness

​What the record never holds

One record, several views

What the record holds

Freshness

What the record never holds