Tell Supercomputer what you want in plain words, a launch teaser, a set of product shots, a short animated scene, and it plans the job, picks models from the catalog and renders the steps. You review what comes back and steer it in the same chat, with no model menus to learn first.
Most AI generators ask you to be the producer: pick a model, tune the settings, write the prompt, judge the output, then choose another model for the next step. An AI creative agent turns that around. You describe the finished result in plain words, and an agent in a chat plans the job, picks models from the catalog, renders each step and shows you what came out. You steer it in the same conversation.
A real creative job is rarely one generation. A campaign teaser might need a still, a cleaned-up cutout of the product, a short animated shot built from that still, and a background track. Done by hand, that is four models, four prompts and a lot of downloading and re-uploading between them. The thinking you actually care about, what the piece should say and how it should feel, gets squeezed between chores.
Supercomputer is the agent in Picasso IA built for that gap. It is a chat in the web app, so there is no key to create and nothing to install. You write what you want, it breaks the work into steps, chooses from a catalog of 502 models, runs them, and reports back with the result and what it did to get there.
It runs on the same models and the same credit system as the toolkit, so a generation costs credits that depend on the model and the settings, and the pricing page is where plans live. New accounts get free credits. The agent does not make generation cheaper or faster by itself. What it removes is the routing: deciding which model fits which step and carrying files between them.
When you send a brief, the agent first reads it for the parts that matter: the subject, the format, the mood, how many pieces you want and what you already have. If something decisive is missing, such as whether the output is a vertical clip or a square post, it asks before spending credits instead of guessing.
Then it drafts a plan in plain language. A brief like a short product teaser from one photo and a moody backdrop might become: clean the photo, generate the backdrop, composite the scene, animate it, add a music bed. For each step it picks from the catalog, perhaps a still model such as nano-banana-pro or seedream-4.5, a cutout model such as remove-background, a video model such as veo-3.1 or kling-v3-video, and a music model such as lyria-3. You see the plan and the choices, so nothing happens out of sight.
Each step renders in turn, and the output of one feeds the next. When a step comes back wrong, you say so in the same chat, in words, and the agent reruns that step rather than the whole job. The conversation is the control surface: a note like warmer light, less clutter, hold the first two shots and redo the third is a valid instruction.
The toolkit and the agent share one catalog and one credit system, but they are different ways of working. The toolkit is the manual workspace: you pick the model yourself, set the options and generate. The agent is the delegated one: you describe the result and it makes the choices. Neither replaces the other.
| Situation | Better fit | Why |
|---|---|---|
| One image from a prompt you already know | Toolkit | You choose the model and settings directly, with no planning step in the way |
| A result that needs several models in a row | Agent | It routes between steps and carries each file forward for you |
| Comparing the same prompt across models | Toolkit | Side by side runs are the point, so you want to pick each model |
| A brief you can describe but not yet spec | Agent | It turns loose words into steps and asks about what is missing |
| Exact control of one setting, like resolution or seed | Toolkit | Manual options are right there, no translating through a chat |
| A set of variations that must feel like one family | Agent | It can hold the brief and apply it across every piece |
A useful habit is to prototype one hard step in the toolkit first, then hand the full job to the agent with the winning wording. Both are open to the same account and the same credits.
This is the loop from blank chat to finished files. It works the same for a single image or a multi-step video job.
Specific briefs make better plans. The three below show the level of detail that helps: a goal, a format, a mood and the material on hand. They are examples of how to write, not promises about a particular output.
A small shop owner might write: I have one photo of a hand-painted ceramic mug. Make three lifestyle images of it on a kitchen table in morning light, then a short vertical clip of steam rising from it. The agent would plan a background step, a placement step for the mug, a still pass, and an image-to-video step, and would show you the stills before the clip.
A musician might write: Make a square cover image for a quiet piano record, dusk colors, no text, then a looping ten second visual for the same mood. The agent would generate the cover, check with you, and extend the same look into motion.
A teacher might write: I need five simple illustrations of the water cycle for ten year olds, same flat style in all five, then upscale them for printing. The agent would lock one style, render the set, and run an upscaling model such as real-esrgan on the ones you approve.
Tip: the best briefs state what to keep as well as what to make. Hold the product shape, keep the logo area empty, same palette in every piece. Constraints give the agent something to check each step against.
An agent is a way to organize generation, not a way around its limits. These are the places to expect friction:
Two rules of use apply everywhere. Use only your own face, voice and likeness, or fictional characters, never a real identifiable person. And if the pieces are for sale, such as product images, keep them truthful about the item. For platform formats and rules, check the current requirements of each platform instead of trusting a number you read in a chat.
It is a chat-based assistant that does the planning part of a creative job. You describe the result you want in plain words, and it breaks the work into steps, picks models from the catalog, renders each step and shows you the outcome. You then correct it by talking to it. Supercomputer is the agent in Picasso IA, and it works with the same models you can use by hand in the toolkit.
The toolkit is the manual workspace: you choose the model, set the options and generate. Supercomputer is the delegated version: you describe the result and it chooses the models and sequence for you. They share the same catalog and the same credit system. Use the toolkit when you already know the model and settings, and the agent when the job needs several steps or you are not sure where to start.
No. Supercomputer is a chat in the web app, so you sign in to your account and write your brief. Keys belong to the API and the MCP server, where a key from your account lets another program call the catalog, and you should treat that key like a password. The agent does not ask for one and you never paste one into the chat.
It picks from the same catalog as the rest of Picasso IA, which holds 502 models across image, video, speech, music, editing and 3D. For a given job that might mean a still model like nano-banana-pro, a video model like veo-3.1, an upscaler like real-esrgan and a music model like lyria-3. You see its choices in the plan and can ask for a different model at any step.
Each generation costs credits, and the amount depends on the model and the settings of that step, so a plan with video costs more than a plan of stills. New accounts get free credits to try things out. The pricing page is the only place worth trusting for plan details, and reading the agent's plan before it runs is the simplest way to see roughly what a job will use.
Yes. The plan is visible before the work starts, and you can answer in plain words, for example use a cheaper model for the draft pass or render the final stills with nano-banana-pro. If you want full control of a single step, you can also run that step yourself in the toolkit and bring the result back into the conversation.
Anything the catalog can generate and chain: images, short videos, edits such as cutouts and upscaling, music beds and voice tracks. The agent adds the planning on top, so jobs with several steps are its strength. It cannot create a result that no catalog model can produce, and it cannot guarantee that the first render will be the one you keep.
Only your own, or fictional characters. Do not use the face, voice or likeness of a real identifiable person who has not agreed to it, whether the request comes through the agent or by hand. If you are building something for sale, keep the pictures truthful about the product, and check the current rules of the platform where it will appear.
Say so in the same chat. Describe what to change in plain words, and ask the agent to redo only that step, so the parts you liked stay as they are. If the same problem keeps coming back, the brief is probably missing a constraint or the model is a poor fit, and asking for a different model usually fixes it. Rerun steps spend credits like any generation.
Everything the agent coordinates is already in the catalog, so you can test the models and the wording in the toolkit now, one step at a time. Write the prompt you would give the agent, run it on a model, and see how the result reads. That practice carries over directly to writing good briefs. If you have questions, the support page is the place to ask.
Ready to test the wording? Open the toolkit and run your first step by hand, or write to support with any question.
Choose your language
