Describe the demo. Get the demo.
Riff Agent reads your changelog, writes the script, drives your staging environment to record the flow, cuts it, voices it, brands it, and hands you a finished video to review. You approve or you redirect. Available on Scale and Enterprise.
- ~9 min
- Prompt to reviewable cut
- Every release
- Triggered from CI
- Human
- Approval always required
- Scale+
- Plan availability
It reads your context
Changelog, Linear issue, PR description, existing docs and your previous videos — so the script sounds like your team wrote it.
It drives your app
Against a staging environment and a demo tenant you designate, with credentials scoped to exactly that.
It never publishes alone
The agent produces drafts. A person approves before anything reaches a customer. That is not configurable.
One sentence, or a whole release
Ask for a specific demo in plain language, or connect it to your release process and let it propose one for every shipped feature. It drafts an outline first — the beats, the angle, the target length and audience — and waits for your nod before recording anything. Most teams spend two minutes on the outline and nothing on the rest.
- Natural-language brief, or triggered by a Linear or GitHub event
- Outline proposed for approval before any recording happens
- Audience, length, tone and brand kit specified up front
- Reuses your glossary, brand kit, voice and presenter
Driving a real application, carefully
You designate a staging URL and a demo tenant, and provide scoped credentials that the agent uses in an isolated browser. It navigates, waits for the interface to settle, retries when something is slow, and records the whole session with the same interaction metadata as a human recording. It will not touch anything outside the environment you designated.
- Isolated browser session with scoped, rotatable credentials
- Environment allowlist — staging only unless you say otherwise
- Self-correcting navigation with retry on slow interfaces
- Full session recording and action log for audit
A draft, and a person who decides
Output lands in your review queue as a normal project: transcript editable, keyframes adjustable, everything overridable. Redirect it in a sentence and it re-cuts. Nothing the agent produces can publish to a customer-facing surface without a human approval, and every agent run is logged with the brief, the actions taken and the approver.
- Lands in the review queue, never live
- Redirect in plain language and it re-cuts
- Human approval mandatory for external publishing
- Full audit: brief, actions, output, approver, timestamp
The technical detail
- Availability
- Scale and Enterprise plans
- Triggers
- Manual brief, Linear issue state, GitHub release, API call, schedule
- Environments
- Explicit allowlist; staging by default
- Credentials
- Scoped, encrypted, rotatable; never reused outside the allowlist
- Typical run
- Around 9 minutes from approved outline to reviewable cut
Security signed off in a week. Private rendering in our own VPC, no training on our frames, SOC 2 report on request. That is not a sentence I get to write about most AI vendors.
About riff agent
Only if you explicitly allowlist a production URL, and we would advise against it. The default configuration permits staging environments only. Whatever you allow, the agent uses credentials you scope and can rotate, in an isolated session, with every action logged.
Tell it what is wrong in a sentence and it re-cuts, often without re-recording. If the flow itself was wrong, it re-records that segment. Because output always lands in review rather than going live, a wrong draft costs you a minute of reading, not a retraction.
It replaces the part of their job that is mechanical — the third variant for a different vertical, the re-record after a UI change, the fourteen release videos nobody had time for. The judgement about what is worth making, and whether the result is good, stays with a person. We designed the approval gate to be non-optional for that reason.
Pairs well with
In production at