Video generation in chat
Describe the clip in the conversation. Approve the spend once. The finished video plays where you asked for it — scrub, mute, fullscreen, download — and the file is stored at your own address.
Approval before every render · clips stored in your workspace · 14-day free trial
23
Video models on the key already in place
1
Approvals needed before a render bills
~90s
What the waiting card tells you to expect
Your workspace
Where the finished clip is stored
Four steps. You are in control at the second one.
Describe it in the conversation the way you would describe it to a person. The agent fills in the prompt, the aspect and the length from what you said. Nothing runs yet and nothing is charged yet.
prompt · aspect · length · read from your message
The render pauses with Approve and Deny in the thread. Approve and the turn resumes and actually runs. Deny and the request stops there. This happens on every video call, not just the first.
approval on every call · no silent spend
The request goes to the video provider, and the card polls for it behind a bounded wait. If the provider is still working when that budget runs out, the card says the render is still going rather than claiming a success that has not happened.
submit → poll → bounded wait → no fake success
The finished bytes are pulled down and written into your workspace storage, then played back inline from your own address. Edit and Variation are right there if the first take is not the one.
stored in your workspace · inline player · edit or vary
One conversation. No second tool to open.
The default path reaches 23 video models on the key the platform already holds — Google Veo 3.1 Lite unless you name another. Ask for a different model in the message and the next render uses it. Nothing to sign up for, nothing to paste in.
Every video call pauses with Approve and Deny in the thread, every time, because every provider bills real money. Deny and nothing is charged. On the web, only a workspace owner can spend on a render — a visitor talking to your agent cannot.
While the render runs you see a card with your prompt, the aspect and length you asked for, and a clock counting up. The progress bar stops short of full on purpose: the exact time left is not known, so no number is invented to fill the gap.
The finished clip appears in the conversation as a player, not a link — scrub, mute, fullscreen, download. Seeking works because the file is served with range requests, so you can drag to the middle without waiting for the whole thing.
Under a finished clip: Edit to change the prompt, Variation for another take on the same idea, Try again after a failure. Each button writes the next message for you, with the length and aspect already filled in.
The finished file is copied into your workspace storage and served from your own path under a random, unguessable name. Provider download links expire within hours. Your copy does not, so a clip you generated last month still plays.
| Feature | ONE | Standalone video tools | Build on provider APIs |
|---|---|---|---|
| Ask for a clip inside the conversation | you build it | ||
| Approval prompt before every billed render | varies | you build it | |
| Finished clip stored under your own address | you build it | ||
| Inline player with seeking and download | in their app | you build it | |
| Switch model per request, one credential | |||
| Same agent also writes, sells and books | you build it | ||
| Full control of the provider contractHonesty: calling the video APIs directly gives you every knob. It also means building the approval step, the storage and the player yourself. |
Every plan can generate video. Renders draw on the same credits everything else does.
Video generation included
Same video tools, more room to run
Dedicated support · SLA
Video generation lives inside the conversation your agent already runs. Nothing renders until you say so, and the finished clip is yours, at your own address.
Approval before every render · no credit card to start