For creators and agencies who like Luma's agent and its Ray and Uni models, but want the work as a graph of nodes they can open and change: Luma's models beside other providers, their own code, their own machines, and a timeline that hands the cut to an editor.
Luma Agents is Luma AI's browser workspace, launched in March 2026. A creative agent plans and generates image, video, audio and text on an infinite board, keeping the project's context from concept to delivery, with Luma's own Ray video and Uni image models beside third-party models. Dream Machine, the earlier app it grew out of, is still open to existing users.
Kaitoi also has an agent, Penny, that takes a brief through concepts, storyboard shots and the timeline, and Luma's Uni-1, Photon and Ray models run as Kaitoi nodes. The difference is that the graphs Penny builds are made of nodes you can open and change, beside your own Python steps and your own machines. Kaitoi is free to start, with no subscription, and hosted models cost the provider's own price with no markup.
Checked against each product's own documentation on .
One agent for the whole project. The Luma Agent plans in Research Mode, writing briefs and shot lists on the board, then generates image, video, audio and text in Create Mode, keeping the project's context from concept to delivery. Scenes turns a brief into an editable storyboard that you approve before the video renders, and the agent can turn a plain description into a reusable Skill.
Its own models, Ray and Uni. Ray, its video model, makes video from text or images, restyles existing footage with Modify Video, reframes it, and can generate native HDR with 16-bit EXR frames. Uni, its image model, generates and edits images and powers Layers, which splits an image into editable layers.
Many other models beside them. Video models such as Veo, Kling and Seedance, image models such as Nano Banana, GPT Image and Seedream, and ElevenLabs for speech, music and sound effects sit in the same workspace, with lip sync too. The agent picks a model for each request, or you open a panel and choose one yourself.
Boards for the whole project. Work lives on infinite boards where assets, notes and frames sit side by side, and versions branch beside each other instead of overwriting. Boards are shared by link or email with viewer or editor access, and a team can work on one in real time, with the agent alongside.
Ready for agencies and enterprise teams. Luma names agencies and brands such as Publicis Groupe and Dentsu among its users. It offers SSO and access controls for teams, its enterprise and API terms promise not to train on your content, Enterprise adds custom fine-tuning, and Ad Variants turns one approved ad into five placement sizes in several languages.
Open to developers and other tools. A REST API serves the Uni and Ray models, with official SDKs for Python, TypeScript and Go and a CLI. A remote MCP server lets Claude, ChatGPT and other assistants run generations and ready-made mini-apps on your Luma credits, and connectors bring files in from Google Drive, Dropbox, Frame.io, Box and Airtable and save work back.
One agent for the whole project. Penny builds and edits graphs, writes node code, lays out storyboard frames, places and splits timeline clips and arranges Lab boards, all in one project. With AdMuse in Lab it takes a brief to references, concepts, scripts and campaign directions, then chains storyboard shots, video Skills and the timeline. Skills, including ones you save, give it reusable actions such as generating an image, producing audio or running a saved workflow, and Playbooks are written instructions it follows. There is no fixed pipeline, and Penny has no approval step before generating: Studio shows a credit estimate before a node or graph runs, and Penny gives one when asked. The difference is that the graphs Penny builds are made of nodes you can open and change.
Luma's own models, as nodes. Luma's Uni-1, Photon and Ray models run as Kaitoi nodes on Kaitoi credits: Uni-1 image generation, Photon image modify, Ray video generation including from an image, and Ray video modify, edit and reframe. Ray's HDR models run as nodes too, returning video rather than Ray's EXR frames. Layer-splitting nodes break an image into layers, and HDR cuts render as 10-bit HEVC or ProRes 4444 XQ. Runway's Ruby node turns an SDR clip into HDR video or an OpenEXR frame sequence, ACEScg included, and other nodes write single EXR images and convert colour spaces. The timeline itself renders no image sequence. Ray and Uni sit in the same graph as other providers' models and your own steps.
Many other models beside them. Veo, Kling and Seedance video models, Nano Banana, GPT Image and Seedream image models, ElevenLabs speech, sound effects and music, and lip-sync models from Sync and Hedra run as nodes, among more than 1,400 generative models as of September 2026. They run on Kaitoi credits at the provider's own price with no markup or, for providers that accept them, on your own keys. Penny picks models as it builds a graph, or you add any model yourself, and when a model has no node yet, Penny can make one from the New Models panel.
Boards beside the conversation. Lab is a canvas of references and results beside a conversation with Penny, and Penny can create boards and arrange what is on them. The project keeps the graph, with backups and templates to return to an earlier version. On a team Studio subscription, team workspaces have roles, shared projects and pooled credits, and share links, on every plan, are read-only with optional cloning. Edits do not sync live: when a teammate saves, Studio offers to load the newer version.
For agencies and enterprise teams. Kaitoi never trains models on your work, on any plan. On a team Studio subscription, team workspaces have roles that control what each member can do, pooled credits and weekly caps per member. Kaitoi has no single sign-on (SSO) yet. A FLUX.2 trainer node trains LoRA weights from your images, AdMuse develops campaign concepts and scripts with Penny, and timeline output can be 9:16, 1:1 or a custom size, with a reframe node to change a clip's aspect with AI. There is no one-click ad-variant tool.
Open to developers and other tools. Kaitoi's MCP server, on every plan, lets compatible AI clients build graphs, run approved work and inspect results, and on Studio write nodes. A REST API runs graphs from your own code, a saved template becomes an API endpoint whose results are polled or streamed, and nodes can use your connected accounts: built-in nodes load video from and upload to Google Drive, though there are none for Dropbox, Frame.io, Box or Airtable. There is no published SDK for the REST API. Penny also calls outside MCP servers, and Kaitoi Code, the terminal agent in Kaitoi CLI, uses Kaitoi workflows from a code project.
Beyond that, any step can be your own code. Nodes are Python with a declared interface, Penny writes the ones you are missing on every plan, and editing node code by hand needs Studio or Enterprise. A node can carry its own HTML, CSS and JavaScript interface, WebAssembly included.
Your hardware can join. Kaitoi CLI connects your own computer or a GPU machine elsewhere, so ComfyUI, Ollama, Blender and Runtime recipes, which Kaitoi Code writes from a Hugging Face model or a GitHub repository, run in the same graph. On Studio, an exported ComfyUI workflow can become one node.
The cut stays in the project. A storyboard feeds a multi-track timeline that renders MP4, WebM or MOV in ProRes and exports or imports OpenTimelineIO with the clip media, for finishing in another editor.
A workflow that works becomes a tool. Publish it as a Station, a simple form or a custom app where each visitor pays for their own runs, listed if you like in a reviewed gallery at kaitoi.io/stations, and share nodes and templates through Community. A Station that runs your own code, Penny-written nodes included, is invite only.
| Question | Luma Agents | Kaitoi |
|---|---|---|
| Agent An AI that plans and carries out the work. | The Luma Agent plans in Research Mode, writing briefs and shot lists on the board, and generates image, video, audio and text in Create Mode, keeping the project's context from concept to delivery. It can turn a plain description into a reusable Skill, and Scenes waits for your approval before it renders. | Penny builds and edits graphs, writes nodes, and edits storyboard frames, timeline clips and Lab boards in one project. With AdMuse it takes a brief to references, concepts, scripts and campaign directions, then chains storyboard shots, video Skills and the timeline. Skills, including ones you save, generate images and audio or run a saved workflow, and Playbooks you write shape it. Studio shows a credit estimate before a run and Penny gives one when asked; Penny has no approval step before generating. |
| Luma's own models Luma's Ray video and Uni image models. | Ray makes video from text or images, restyles footage with Modify Video up to 20 seconds at 1080p, reframes it, and can generate native HDR with 16-bit EXR frames. Uni generates and edits images and powers Layers. | Luma's Uni-1, Photon and Ray models run as Kaitoi nodes on Kaitoi credits: Uni-1 image generation, Photon image modify, Ray video generation including from an image, and Ray video modify, edit and reframe. Ray's HDR models run as nodes too, returning video rather than Ray's EXR frames. Layer-splitting nodes break an image into layers, and HDR cuts render as 10-bit HEVC or ProRes 4444 XQ. Runway's Ruby node turns an SDR clip into HDR video or an OpenEXR frame sequence, ACEScg included, and other nodes write single EXR images and convert colour spaces. The timeline itself renders no image sequence. |
| Other models Third-party models you can reach in the same place. | Video models such as Veo, Kling and Seedance, image models such as Nano Banana, GPT Image and Seedream, and ElevenLabs for speech, music and sound effects, with lip sync. The agent picks a model for each request, or you choose one in a panel. | Veo, Kling and Seedance video models, Nano Banana, GPT Image and Seedream image models, ElevenLabs speech, sound effects and music, and lip-sync models from Sync and Hedra run as nodes, among more than 1,400 generative models as of September 2026, on credits or your own keys. Penny picks models as it builds a graph, or you add any model yourself, and Penny can make a node for a model that has none from the New Models panel. |
| Editing images and video Changing and finishing media you already have. | Layers splits images into editable layers. Images are edited by instruction, reframed and cut out, and images upscale to 8K and video to 4K, with SeedVR and Topaz among the upscalers. | Layer splitting, background removal, inpaint, outpaint and erase run as nodes, and Nano Banana, GPT Image, Seedream, FLUX.2 and Qwen Image edit nodes change an image from a written instruction. Image upscalers from Magnific, Recraft, Bria and SeedVR, and Topaz and FLUX video upscaling, run as nodes, and a Ray reframe node changes a clip's aspect with AI. |
| Storyboard and cut Planning shots and assembling them into a finished piece. | Scenes turns a brief into an editable storyboard of keyframes, approved before the video renders, and is built for short-form sequencing. Clips are trimmed and split on the board and stitched by the agent, and export as MP4, with EXR alongside for HDR. | A storyboard of scenes and shots, which Penny can edit, feeds a multi-track timeline where Penny places and splits clips. The cut renders MP4, WebM or MOV in ProRes, and exports and imports OpenTimelineIO with the clip media for another editor. Studio shows a credit estimate before a run, without an approval step. Runway's Ruby node turns an SDR clip into HDR video or an OpenEXR frame sequence, ACEScg included, and other nodes write single EXR images and convert colour spaces. The timeline itself renders no image sequence. |
| Board and teamwork Where the work sits, and working on it with others. | Infinite boards hold assets, notes and frames, with versions side by side. Boards share by link or email with viewer or editor access, a team can work on one in real time, and any Luma account can add members with admin or member roles and a shared credit pool. | Lab puts references and results on a canvas beside the conversation with Penny, and the project keeps the graph, with backups and templates to return to an earlier version. Team workspaces with roles, shared projects and pooled credits need a team Studio subscription, and share links are read-only with optional cloning. Edits do not sync live: when a teammate saves, Studio offers to load the newer version. |
| Training and enterprise terms What happens to your content, and controls for larger teams. | Luma's standard terms let it use content to train models, though Luma says user interaction data was not used for its current Uni and Ray models. Enterprise and API terms promise no training. SSO and access controls are offered for teams, and Enterprise adds custom fine-tuning. | Kaitoi never trains models on your work, on any plan; a provider called by a node handles data under its own terms. On a team Studio subscription, team workspaces have roles that control what each member can do, with weekly spending caps; Kaitoi has no single sign-on (SSO) yet. A FLUX.2 trainer node trains LoRA weights from your images, on every plan. |
| Advertising Tools for campaigns and ad production. | Research Mode gathers trends, competitor examples and references. Ad Variants turns one approved ad into five placement sizes in several languages, and one-click mini-apps include relighting and background swap. | AdMuse works with Penny in Lab to take a brief to references, concepts, scripts and campaign directions, keeping the sources with each direction. Timeline output can be 9:16, 1:1 or a custom size, and a reframe node changes a clip's aspect with AI. Background removal and instruction edits run as nodes, and a MiniMax model relights video. There is no one-click ad-variant tool and no image relighting node. |
| Developers and other tools Reaching the product from code, assistants and other services. | A REST API for the Uni and Ray models, with Python, TypeScript and Go SDKs and a CLI. A remote MCP server lets Claude, ChatGPT and other assistants run generations and mini-apps on your Luma credits. Connectors bring files from Google Drive, Dropbox, Frame.io, Box and Airtable and save work back. | Kaitoi's MCP server, on every plan, lets compatible AI clients build graphs, run approved work and inspect results, and on Studio write nodes. A REST API runs graphs from your own code, a saved template becomes an API endpoint whose results are polled or streamed, Kaitoi CLI brings Kaitoi Code to your terminal, Penny calls outside MCP servers, and nodes can use your connected accounts: built-in nodes load video from and upload to Google Drive, though there are none for Dropbox, Frame.io, Box or Airtable. There is no published SDK for the REST API. |
| Reusing and sharing workflows Running a workflow again, and letting others run it. | Skills save a workflow in plain language; they can be forked, chained and shared by link, and recipients install them in their own boards. The Discover tab lists ready-made tools such as Ad Variants. | Penny's Skills can run a workflow you saved, Playbooks hold written instructions, a saved template starts new projects and shares by link, and Create subgraph folds steps into one reusable node. Nodes and templates publish to Community, and a Station turns a workflow into a form or a custom app where each visitor pays for their own runs, listed if you like in a reviewed gallery at kaitoi.io/stations; a Station that runs your own code is invite only. |
| Where it runs What you use it on, and whose hardware does the work. | A hosted service in the browser, with generation in Luma's cloud. An iPhone app, still listed as Luma Dream Machine, and mobile browsers. | Studio is hosted in the browser and works on mobile and touch screens, with no native app. Kaitoi CLI connects your own computer or a GPU machine elsewhere, so ComfyUI, Ollama, Blender and Runtime recipes that Kaitoi Code writes from a Hugging Face model or GitHub repository run in the same graph. |
Each cell comes from the product's own website, documentation or code repository. "Not publicly documented" means we found no official page that says so, not that the feature is missing. Plans and features change; check the official pages before you decide.
A simple conversation evolves into a complete ad as Penny moves through concept exploration, visual development, storyboard, shot generation, timeline, and rough cut assembly inside Kaitoi.
This film, from our Kaitoi Signals series, follows one ad from the first message to a rough cut.
Read more in the documentation:
Plus, Pro and Ultra plans, billed monthly or yearly, include credits that reset each billing cycle, and credits for failed generations are returned. There are free trial credits, commercial use comes with the paid plans, Enterprise is sold through sales, and the API is billed pay-as-you-go in credits.
Free to start, with no subscription. Credits pay for hosted models, including Penny's and Luma's, at the provider's own price with no markup, and for paid Community nodes; steps that call neither cost nothing. Your own provider keys work on every plan and are billed by your provider. Studio adds included credits, team seats and hand-edited code.
Not quite. Luma says Luma Agents grew out of Dream Machine, its earlier app from 2024, and is not just a new name for it. Dream Machine stays open to existing users, and moving boards across is a manual download and upload.
Yes. Luma's Uni-1, Photon and Ray models run as Kaitoi nodes on Kaitoi credits, next to models from other providers. Luma Agents itself, Luma's agent and its boards, is not part of Kaitoi.
Luma's standard terms allow it to use content for training, its enterprise and API terms promise no training, and Luma says user interaction data was not used for its current Uni and Ray models. Kaitoi never trains models on your work, on any plan; when a Kaitoi node calls Luma or another provider, that provider handles the data under its own terms.
Yes, on every plan. Compatible AI clients can build graphs, run approved work and inspect results through it, and Penny can also call outside MCP servers.