One API for every AI video use case
Send one prompt. Get back a finished, minutes-long video — plus an editable project you can regenerate scene by scene.
Running as an A/B test inside Vivideo, our own AI video app.
POST https://api.longvid.ai/v1/videos
Authorization: Bearer YOUR_API_KEY
{
"prompt": "A 75-second cinematic short film: a lighthouse keeper on a stormy Norwegian coast finds a message in a bottle from his younger self."
}The only thing you send is the prompt.
A 75-second cinematic short film: a lighthouse keeper on a stormy Norwegian coast finds a message in a bottle from his younger self.
Every example is a single prompt — the exact object sent to the API. Step through with the arrows.
Models make clips. Finished videos take a pipeline.
A model gives you a few seconds at a time — and can't shoot real places at all. Shipping a real video means building the whole pipeline, then maintaining it forever.
Build the pipeline yourself
- 01Write the script and split it into scenes
- 02Wire up a model API for every scene
- 03Find real footage models can't make — like Istanbul
- 04Stitch voice, music, captions; run QC and retries
One Longvid API call
- Send one prompt, get the finished video back
- Real footage pulled in automatically where models fall short
- Every scene comes back editable — regenerate one at a time
- We route each scene to the best model as new ones ship
- You never build or maintain a pipeline again
We built it for Vivideo, our own AI video app, because the third-party API we used before kept failing us.
Any kind of video. One API.
Cinematic films, faceless explainers, enterprise training, UGC ads — plus repurposing existing videos and turning PDFs, files or URLs into video. It all goes through one call and comes back finished.
longvid
one API call
What one API call gives you
One prompt in, a finished film out — and the pieces to change it.
Any format
Every kind of video, from one prompt
UGC ads, explainers, a brand spokesperson, a weekly show from a blog, a localized cut for another market. You describe the outcome; Longvid writes the script, casts the voices and renders the shots.
- Long-form, short-form, vertical or wide
- Voiceover and music scored to the cut
- Burned-in captions in the language you ask for

Real footage
Real footage where the models fall short
Some shots a model just can't make — real footage of Istanbul for a travel video, a specific product, an actual place. Longvid pulls that in and cuts it together with the generated scenes, so the whole film holds up.
- Generated and real footage in one timeline
- The right source chosen per shot, automatically

Editable
Not one opaque file. Every scene, editable.
The video comes back as a storyboard of scenes, each one its own piece. Regenerate a single scene, or change a delivered video by telling the director what you want in plain words — the rest stays exactly as it was.
- Regenerate one scene without touching the others
- Edit in words: “re-make scene 3 at dusk”
- Open the project at app.longvid.ai/studio for new takes

Long-form
Minutes-long videos that stay coherent
The hard part isn't a five-second clip — it's a five to ten minute video that holds together across dozens of scenes. That coherence is what we tune for, and it's the wedge everyone else is worst at.
- Consistent characters, look and pacing across scenes
- Tuned on real users before any customer sees a change

Longvid vs. the alternatives
Avatar tools and faceless generators each do a slice. Longvid is the whole pipeline — many models, real footage and long-form coherence — behind one call.
| Capability | LongvidOne API call | HeyGen v3 AgentAvatar-first | VideoGen APIFaceless-first | Build it yourselfIn-house pipeline |
|---|---|---|---|---|
| What you send | Prompt · files, images, video, URLs | Prompt · files, PDF, video | A full script (verbatim) | Prompts + code per scene |
| Primary output | Long-form multi-scene film | Avatar / talking-head | Faceless short & mid-form | Whatever you assemble |
| Range of video types | Cinematic, faceless, UGC, enterprise, repurpose & clips | Avatar-first | Faceless-first | Anything — if you build it |
| Best model picked per scene | ? | You build it | ||
| Real footage where models can't | You source it | |||
| Built for minutes-long coherence | ? | ? | You build it | |
| Edit one scene after delivery | You build it |
Based on public docs and live API tests of HeyGen and VideoGen, Sept 2026. Capabilities and pricing change — check each vendor for the latest.
Tested head to head inside Vivideo
Longvid has an MVP running as an A/B test inside Vivideo. We test three approaches head to head on real users.
Our pipeline
Our previous default
After around 2 weeks, Longvid users get and keep more finished videos:
The download gap is the strongest signal and is statistically significant (z ≈ 3.7).
VideoGen's videos were so irrelevant that we had to stop testing it. Check the examples yourself.
One call runs the entire studio
POST a prompt to /v1/videos and the pipeline handles the whole production — planning, assets, rendering, editing, QC and export. Poll the video until its status is terminal, or register a webhook and forget the polling.
Story, shot list and storyboard.
Voices, characters and places.
Every scene is rendered.
Cut, music and captions.
Quality check.
Final encode.
Ship it, or approve the plan first
By default the video renders straight through. Set approval: manual and it stops at the storyboard so you can approve, decline, or pick a tier before anything is rendered.
Know exactly where it is
The video object carries status, stage and progress.pct. When it's done, you get a signed video_url, thumbnail, duration and scene count.
Typical wall time: a few minutes per minute of video.
A prompt in, a finished video out
The whole request is a prompt plus a few optional knobs — duration, tier, aspect ratio. Send it, then read the finished video object back.
{
"prompt": "A 60-second cinematic reveal
of an electric grand tourer",
"duration_sec": 60,
"tier": "premium",
"aspect_ratio": "16:9"
}- One endpoint to create, one to poll, one webhook when it is done
- Same account and credits — no separate plan to enable the API
- Idempotency keys, per-video webhooks and signed URLs built in
Then poll or wait. Read GET /v1/videos/{id} no more than every 10 seconds, or register a webhook and get video.completed when the file is ready.
Call it however you build
The same pipeline, everywhere you work — REST, MCP, CLI, agent skills and SDKs. Every one is documented in one place.
REST API
One HTTP call to create, one to poll, one webhook when it lands. JSON in, video out.
POST /v1/videosMCP server
Drop Longvid into Claude, Cursor or any MCP client and generate videos straight from your agent.
mcp.longvid.aiCLI
Kick off and download videos from your terminal or CI — perfect for scripts and pipelines.
$ longvid createAgent skills
Ready-made skills that teach coding agents how to brief Longvid and ship a finished video.
longvid/skillsSDKs
Typed clients for TypeScript and Python so you can call the pipeline in a couple of lines.
npm · pipRead the docs
Guides, references and quickstarts for every surface live at developers.longvid.ai.
Model-agnostic on purpose
A new model ships almost every week. We route each scene to the model that wins for that shot, and swap in new ones as they land — so you never rebuild your pipeline again.
Seedance 2.5
ByteDance
Kling O3
Kuaishou
Gemini Omni
Google DeepMind
ElevenLabs
Voice & audio
Video is routed per scene across the leading models — Seedance, Kling, Wan, Gemini, Vidu, PixVerse and MiniMax; avatars and lip-sync via HeyGen, OmniHuman and more; voice from ElevenLabs. Where models can't produce a shot, we pull in real footage from Pexels, Pixabay, Coverr, Storyblocks and NASA.
The whole pipeline runs on Python on AWS, and LLMs write the script, plan the scenes and orchestrate the run. When a better model appears, we route to it — the API you call stays the same.
Pay as you go. Nothing until you generate.
It's an API — there's no plan to buy. You pay only for the video you make. Standard quality is $12 for one minute of finished video.
No plan, no commitment. You're billed only for the video you generate.
One minute of finished video costs $12 in standard quality.
- No subscription, no seats, no minimums — $0 to start
- Charged per second of delivered video, rounded up
- Requested length is prepaid, then settled to the real length
- New accounts start with welcome credits to try it free
Estimate your monthly cost
pay as you goEstimated monthly cost
30 min · standard
$360
Charged per second of delivered video, rounded up. Need more than 600 min/month? longvid@vivideo.ai
Questions
It turns one prompt into a finished, minutes-long video. It writes the script, splits it into scenes, routes each scene to the best model, pulls in real footage where models fall short, and stitches it with voice, music and captions. You get the video plus an editable project.
Today, AI video apps that would otherwise build and maintain this pipeline themselves. Over time, any developer who needs video from an API.
Yes. Every video comes back as a storyboard of scenes you can regenerate one at a time. You can also edit a delivered video by telling the director what to change in plain words, and open the project in the studio for new takes and another export.
A model API gives you a clip of a few seconds. Longvid is the pipeline above the models: planning, scene splitting, per-scene model routing, real footage, voice, editing, QC, retries, queues, webhooks and billing — so you make one call instead of building and maintaining all of that.
Video generation is routed per scene across the leading models — Seedance, Kling, Wan, Gemini, Vidu, PixVerse and MiniMax — with avatars and lip-sync (HeyGen, OmniHuman), voice from ElevenLabs, and real stock footage where models fall short. We stay model-agnostic on purpose and swap in new models where they win, so the API you call stays the same.
It's pay as you go — $0 to start, no subscription, no seats, no minimums. You're billed only for the video you generate: standard quality is $12 per minute of finished video, charged per second and rounded up. New accounts start with welcome credits to try it free.
Longvid has an MVP running as an A/B test inside Vivideo, our own AI video app. We test three approaches head to head: Longvid (1 group), HeyGen's agent API (14 groups, as it was our default) and VideoGen's API (1 group). VideoGen's videos were so irrelevant that we had to stop testing it. After around 2 weeks, Longvid users get and keep more finished videos: they complete 28% more and download 50% more. The download gap is the strongest signal and is statistically significant (z ≈ 3.7).
Make your first video
Create a key in Settings, paste it into the curl, and your first video is queued.