Decart
Decart is a frontier AI lab and developer platform whose real-time world models edit live video, generate interactive simulated environments and optimize inference across major AI chips, all reachable through a single usage-priced API.
What is Decart?
Decart describes itself as a frontier AI lab building real-time world models, and sells access to them through a developer platform. Three pieces sit on top of one another. DOS, the Decart Optimization Stack, is the performance layer: it is presented as unlocking the full throughput of every major chip, built across NVIDIA, AWS Trainium and Google TPU so that inference and training run faster, cheaper and at higher utilization. DOS 2.0 was announced alongside the company's 300 million dollar round. Lucy and Oasis are the models that run on it.
Lucy edits video as it plays. Billed as the only world editing model at production scale, it swaps or adds characters, inserts, replaces and removes objects, rewrites backgrounds, applies new styles, changes attributes and generates visual effects, all while the content is being created, streamed or watched. The product page targets shopping, advertising, live streaming, social platforms and gaming, and puts hard numbers on the claim: sub-40 millisecond latency and 100 FPS of continuous generation in the banner, with the same page's FAQ quoting under 200 milliseconds at 22 FPS in 512x768. Its siblings on the API are Lucy VTON for virtual try-on, Lucy Restyle for style transfer and Lucy Image for still editing.
Oasis generates worlds rather than editing them. Oasis 3 produces controllable, physically plausible environments in real time, action-conditioned so that each next world state is generated autoregressively from the policy's latest move, with synchronized multi-camera output from one coherent state and end-to-end latency under 200 milliseconds. It is aimed first at autonomous vehicles, then at robotics, drones, manufacturing and maritime, and is offered as the first interactive world model reachable through an API.
Developers get three entry points: a Realtime API for live camera and video, a Queue API for asynchronous video jobs, and a Process API for synchronous image editing. Official SDKs cover JavaScript, Python, Swift and Android, with Vercel AI SDK, TanStack AI and LangChain.js integrations. Pricing is metered per second or per generation, with no subscription. Decart ships no text, chat or audio model of its own.
What it does
- Edit a live video stream from a text prompt, a reference image, or both
- Swap or add a character in a running feed, and change attributes on the fly
- Add, replace or remove objects and transform the background in real time
- Apply a new visual style to a stream and generate physically aware effects
- Place products on the filmed person for real-time virtual try-on
- Generate an interactive 3D environment a robot policy can act inside
- Render synchronized multi-camera views of a single coherent world state
When to use Decart / When not to
A quick filter to help you decide if Decart is the right fit.
When to use Decart
- Product engineers embedding generative video into a live product through an API
- Streaming platforms and creators who want the picture transformed while it is being broadcast
- Retail and e-commerce teams building virtual try-on and in-context product placement
- Advertising and brand teams turning already-playing video into addressable ad inventory
- Robotics and autonomous-vehicle teams that need controllable simulated environments for closed-loop training
When not to use Decart
- Editors who want a timeline-based video suite with no code involved
- Anyone looking for a text assistant: Decart ships no chat, writing or reasoning model
- Buyers who need a fixed monthly fee, since everything is metered per second or per generation
- Teams needing native audio or music generation, which the platform does not produce itself
- Under-13s everywhere and under-16s in the European Economic Area, who are excluded by the terms
How to use Decart
A typical end-to-end flow, from setup to results.
- Try the public demos first at lucy.decart.ai and oasis3-preview.decart.ai, which need no account
- For an enterprise evaluation, fill in the Book a Demo form and pick your headcount band, from 1-10 to 1000+
- To build, create an account on platform.decart.ai, which comes with free evaluation credits
- Generate an API key from the API Keys section of the platform
- Install a client: npm install @decartai/sdk, or the Python, Swift or Android SDK
- Pick the right API: Realtime for live camera work, Queue for asynchronous video jobs, Process for synchronous image edits
- Never ship the permanent key to a browser; issue short-lived client tokens or route calls through a server-side proxy
- Open the realtime session over WebRTC, passing the model, the text prompt and any reference image
- Follow the streaming best-practices page for camera permissions, stream constraints and unsupported environments
- For Oasis 3, install the separate decart-oasis Python SDK, set DECART_API_KEY, prompt the environment and start sending control signals
Pros & Cons
Pros
- Latency is the point of the product, and it is quantified: under 40 ms claimed for Lucy, under 200 ms end-to-end for Oasis 3
- Pricing is fully public and metered, with no subscription and no minimum spend
- New accounts get free credits, so the API can be evaluated before any payment
- Output ownership is assigned to the customer, with commercial use explicitly permitted
- Developer documentation is unusually complete: quickstart, API reference, cookbook, changelog, FAQ and a public status page
- Compliance posture is documented rather than asserted: published DPA with EU SCCs, named sub-processors with countries, SOC 2 Type 2 and ISO 27001/42001 declared
- Oasis 3 is offered as the only interactive world model available through an API, which has no direct equivalent
Cons
- No permanent free plan: past the evaluation credits, every second is billed
- Per-second billing adds up quickly, with a 60-second Oasis 3 session costing 1.20 USD by the vendor's own example
- Without code you get no further than the public demos, since there is no non-technical interface
- Oasis 3 is still labelled Preview, and its enterprise pricing is quoted rather than published
- Oasis 3 requires its own Python SDK instead of the unified client used by every other model
- The terms let Decart use customer Content for platform development and for marketing, with no documented way to opt out
- Performance figures contradict each other on the same page: under 40 ms and 100 FPS in the banner, under 200 ms and 22 FPS in the FAQ
Pricing & Plans
There is no permanent free plan. New accounts receive free credits for evaluation only, after which all usage is metered. The published entry point is 0.01 USD, charged per second for Lucy Restyle 2 and per generation for a 480p image with Lucy Image 2. Volume and enterprise pricing is quoted on request.
- Decart charges a unit price per model
- pay-as-you-go
- with no minimum spend
- Free evaluation credits granted on account creation
- 0.02 USD per second
- 0.02 USD per second
- 0.01 USD per second
- 0.02 USD per second
- enterprise pricing on request
- 0.04 USD per generated second
- 0.01 USD per generated second
- Lucy 2.1
- Lucy VTON 3 and Lucy VTON 2 at 0.04 USD per generated second
- Lucy Clip at 0.15 USD
- 0.01 USD per 480p generation
- 0.02 USD per 720p generation
- Lucy 2.1
- Lucy VTON 3 and Lucy VTON 2 at 0.02 USD per second
- on quotation via contact@decart.ai
Data, GDPR & hosting
A consolidated view of how Decart handles your data.
GDPR overview
Implementation is concrete rather than declarative. The privacy policy lists a GDPR legal basis for each purpose (consent, legitimate interest, performance of a contract) and sets out the full catalogue of rights: access, rectification, erasure, objection, restriction, portability, objection to profiling and withdrawal of consent, exercised at contact@decart.ai. Transfers outside the EEA rely on European Commission adequacy decisions under Article 45, on Standard Contractual Clauses, or on another approved mechanism. A signable Data Processing Agreement is published, incorporating the EU SCCs, naming a data protection officer at dpo@decart.ai and listing every sub-processor with its country in Annex III. Annex II states that SOC 2 Type 2, ISO 27001 and ISO 42001 certifications are maintained, with evidence on request. One gap: no Article 27 representative in the European Union is designated anywhere on the site.
Who owns the data?
The terms draw a clear line. You own your Input, and Decart assigns to you all of its right, title and interest in the Output, so generated frames and images are yours to sell or publish. That grant is conditional on your continued compliance with the terms, and you alone warrant that you had the rights to the material you fed in. Decart keeps a broad licence of its own: it may use Content to develop, improve, provide and maintain the platform, to further its business, to comply with the law and to enforce its rights, and separately for marketing and promotion, including on social media. On the API side the published DPA casts Decart as processor and the developer as controller.
Reuse rights
Output can be reused without asking anyone. The terms state that you may use your Content for any purpose, commercial ones such as sale or publication included, provided you keep to the terms and infringe no third-party right. Two caveats sit next to that permission. Output is not unique: the same or a similar result may be generated for another customer, and results produced for others are not your Content. And Decart's own Materials, meaning the site copy, marks, demo footage and software, remain its property and are licensed for personal, non-commercial viewing only. Feedback and suggestions you send may be implemented freely, with no obligation or compensation. Decart also warns that Output can be factually wrong and recommends human review. The Acceptable Use Policy, published separately on the developer platform, was not reviewed for this record, so the detailed list of prohibited uses is not covered here.
Data retention & training
Hosting summary
No page names a hosting country directly; the answer is in Annex III of the published Data Processing Agreement, which lists every sub-processor with its location. Amazon Web Services and Google Cloud host and process customer data in the United States, alongside Hive for content moderation, and ElevenLabs, OpenAI and Fal for specific features, all United States based. The one exception is Veriff, in Estonia, which handles identity verification. The data importer under the Standard Contractual Clauses is Decart.AI, Inc. in Wilmington, Delaware. Transfers out of the EEA rely on Commission adequacy decisions under Article 45, on Standard Contractual Clauses, or on another approved mechanism. The security annex adds that physical access control is delegated to those cloud providers, that infrastructure processing customer data is physically and logically segregated from infrastructure that does not, and that disk encryption and regular backups are in place. No broad hosting region such as EU or North America is claimed as such, and no data residency choice is offered.
Things to keep in mind
Risks and trade-offs to weigh before adopting Decart.
- Real-time editing makes a doctored live stream indistinguishable from an untouched one, which is a direct route to misinformation and to misrepresenting people on camera
- Character swaps and face insertion from a reference image raise a consent question about the people depicted, and the terms put that responsibility entirely on you
- Decart reserves the right to use your Content to develop the platform and to promote itself, including on social media, and no opt-out is documented
- Output is not unique across customers, so a competitor can legitimately end up with a very similar result
- Billing runs by the second: a realtime session left open keeps costing money with nothing to show for it
- Inserting advertising into footage that looks untouched erodes the line between content and paid placement for the viewer
- A photorealistic simulation is still a simulation; validating an autonomous system inside Oasis cannot replace testing it in the real world
Setup & Integrations
Technical difficulty
Two very different answers. Trying it costs nothing: the public demos run in a browser with no account. Integrating it is a developer job. A first realtime session is a few lines with the official SDK, but the permanent API key must never reach the browser, so short-lived client tokens or a server-side proxy are mandatory, and realtime means WebRTC, with camera permissions, stream constraints and unsupported environments as the documented failure modes. Oasis 3 adds a separate Python SDK. Framework integrations and the cookbook shorten the path.
Deployment
Integrations
Behind Decart
Fundraising
Social
Resources
All the official URLs gathered for verification and reference.
Alternatives
Tools that compete with or complement Decart.
Frequently asked questions
What exactly does Decart sell?
What can Lucy actually change in a video?
What is Oasis 3 for?
How much does it cost?
Is there a free tier?
Can I sell what I generate?
Is my content used to train Decart's models?
Where is the data hosted, and is there a DPA?
How hard is it to get running?
Is there a minimum age?
Should you pick Decart?
Decart has picked one hard problem, latency, and built a whole stack around it. The proposition is coherent from bottom to top: DOS squeezes throughput out of NVIDIA, Trainium and TPU silicon, and Lucy and Oasis are what that headroom buys, one editing video while it plays and the other generating worlds a machine can act inside. The commercial packaging is more mature than the research framing suggests, with public per-second pricing, official SDKs in four languages, a status page, a signable DPA and declared SOC 2 Type 2 and ISO certifications. More than 450 million USD raised and named industrial partnerships back that up.
The reservations are just as clear. This is a developer product: without engineering time you will not get past the public demos, and there is no fixed monthly price to plan around, only meters that keep running. Oasis 3 is still a Preview whose enterprise pricing is quoted rather than published, and it needs its own SDK. The latency numbers on the Lucy page disagree with each other, banner against FAQ, which is an odd thing to leave standing. Legally, the terms let Decart use customer Content to develop the platform and to market it, with no documented opt-out, and the whole structure is anchored in Delaware and New York law with mandatory arbitration and no EU Article 27 representative.
For teams building live video into a product, virtual try-on into a storefront, or simulated environments into a robotics pipeline, Decart is one of the few places where any of that runs fast enough to be interactive. For anyone wanting a finished creative tool, it is the wrong shelf entirely.
- Choosing a selection results in a full page refresh.
- Opens in a new window.