Connect the Picasso IA MCP server to Cursor and ask for the hero banner, the icon set or the empty-state illustration in the same chat that writes your code. The result comes back as a file you can save into the project, from the same catalog and the same credits you use in the toolkit.
You are in Cursor, the layout finally works, and the screen still needs a hero image, an empty-state illustration and a set of icons. The usual routine is to open another tab, generate somewhere else, download, rename and drag the files into the project. With the Picasso IA MCP server connected, you ask for the asset in the same chat that is writing your code, and the result comes back as a file you can save into the project.
Placeholder art has a way of surviving into production. A gray box labeled hero image sits in the layout for weeks because getting a real one means leaving the editor, and leaving the editor means losing the thread of what you were building. Generating inside the editor closes that gap: the assistant already knows the screen, the copy and the mood you described, so the image prompt can borrow all of it instead of starting from a blank text box.
MCP, the Model Context Protocol, is an open protocol that lets an AI assistant call external tools. Cursor is one of the clients that can use it, and the Picasso IA MCP server is one such tool: it gives the assistant access to image generation, and to video, voice and more, from the same catalog that runs in the toolkit. That catalog holds 502 models, 212 of them text-to-image, so you can pick the engine that suits the asset rather than settle for one house style. The all-models page shows what is in it.
Credits work the way they do in the toolkit. Every generation costs credits, the amount depends on the model and the settings, and new accounts get free credits to start. Plans live on the pricing page. No number is quoted here, because the number belongs to the model you choose.
Three parts are involved, and keeping them apart makes the setup easy to reason about. Cursor is the client: it hosts the chat, reads your project and decides when a tool is worth calling. The Picasso IA MCP server is the tool provider: it receives the request, runs the generation on the model you named, and returns the result as a file you can download. Your account sits behind it, because the credits are yours.
Access works with a key you create in your account. Treat it like a password: keep it out of the repository, out of screenshots and out of any configuration file you commit, and create a new one if it ever leaks. The key goes into the server entry in Cursor's MCP settings, not into the chat, so the assistant never needs to paste it anywhere.
One request produces one generation. Ask for an icon and you get an icon. Ask for four variants and the assistant makes the request more than once, which costs credits each time. Generation takes time, longer on some models than on others, so a request is not an instant answer the way a code completion is.
The assistant will pick a model on its own if you let it, but naming one gives you control over cost and look. Image models differ more than people expect: one holds a photographic mood, another keeps flat shapes clean, another edits an existing picture without redrawing the rest. This table is a starting point, not a ranking, and the toolkit is the fast place to test any of these on a prompt before you wire it into a habit.
| Asset | Good starting point | Ask for | Check before shipping |
|---|---|---|---|
| Hero or landing banner | flux-2-pro, seedream-4.5 | Wide frame, quiet space on one side for a headline | How it crops at phone width |
| App icon or small mark | flux-2-pro, ideogram-v3-quality | One centered object, simple shapes, plain background | Whether it still reads at small size |
| Vector icons and logo marks | recraft-v4-svg | Flat shapes, one palette, a consistent stroke | That the SVG opens and edits cleanly |
| Empty-state or onboarding art | seedream-4.5 | A calm scene in your brand colors | Whether the style matches your other screens |
| Edit an existing asset | nano-banana, qwen-image-edit | Change one thing in an uploaded image | That the rest of the picture stayed the same |
Need a transparent background for a sticker, a card or a layered mockup? Generate first, then run the result through bria/remove-background in a second request. The transparent PNG maker page walks through that step in detail.
This is the path from an empty folder to a file in your project. It assumes you have a Picasso IA account and Cursor open on a project that needs pictures.
If you want to try a prompt before you connect anything, the toolkit runs the same models today. The one difference is where you type: there it is a form, here it is the chat.
The assistant writes decent image prompts when you give it a clear brief, and mediocre ones when you say make an image for the home page. These habits are what the work consists of.
Tip: ask the assistant to show you the exact prompt it sent before you generate variants. Reading it takes ten seconds, and it usually reveals why a result missed: a vague word, a missing color or a placement you forgot to mention.
Generating images from the editor saves trips, but it does not remove the steps that still need your eyes. Be clear on the edges before you rely on it.
Text inside images is the first one. Headlines, button labels and logos with letters can come back misspelled or warped, and a model such as recraft-v4-svg or ideogram-v3-quality is no substitute for setting real text in your own design file. Keep words in code or in your design tool and let the model draw the picture around the space.
Consistency is the second. Ten icons generated in ten separate requests will drift in stroke weight and style. A style sentence and reference images in a model that accepts them, flux-2-pro for instance, narrow the drift but do not erase it, and a set that must match perfectly is better refined by hand.
Cost and time are the third. Each request spends credits and takes as long as the model needs, so a loop of twenty tries is a decision, not a free default. Prototype in the toolkit, then settle on a model before you ask the assistant to produce a batch.
Finally, outputs need review. Check anything with a face, a trademark or a claim about a product for what it shows, and check the license and platform rules that apply to the app you are shipping. This page describes a way to produce the picture, not a guarantee that it is ready to publish.
Yes, through MCP. The Picasso IA MCP server gives an MCP-capable client such as Cursor access to image generation, so you ask for an image in the chat and the result comes back as a file you can save into your project. It uses the same models and the same credits as the toolkit, so nothing about the catalog is different from what you see on the website.
MCP, the Model Context Protocol, is an open protocol that lets an AI assistant call external tools. Cursor does not generate pictures by itself, so an MCP server is how it reaches a service that does. Once the server is added in Cursor's MCP settings, the assistant can decide to call it when your request needs an image, and show you what came back in the conversation.
Yes. Access works with a key you create in your account, and you should treat it like a password: never commit it, never paste it into a public chat or a screenshot, and replace it if it leaks. The key ties each request to your credits. In the toolkit you simply work from the web app, signed in to your account.
The same catalog as the toolkit: 502 models in all, 212 of them text-to-image. For app work, the useful groups are general image models such as flux-2-pro and seedream-4.5, vector output with recraft-v4-svg, and edit models like nano-banana and qwen-image-edit. You can compare them on the all-models page and name the one you want in your request.
Each generation costs credits, and how many depends on the model and the settings you choose, so no single number applies. New accounts get free credits to start, which is enough to see whether a model fits your asset. Current plans are listed on the pricing page, and that page is the only place worth trusting for figures. Cheaper models are a sensible choice while you are still iterating on a prompt.
The models and credits are the same. What changes is where you work: instead of filling in a form in the toolkit, you describe the asset in the Cursor chat and the assistant makes the request for you. Many people test a prompt in the toolkit first, because you see every setting at once, and then reuse the wording in the editor once it works.
It can ask a vector model for them. recraft-v4-svg produces SVG files with editable paths, which is the right starting point for flat icons and simple marks. Treat the output as a draft: open it in a vector editor, check that the paths are clean and the shapes are consistent across a set, and fix what a designer would fix before it goes into a product.
Only if you tell the assistant what the style is. Name your palette, describe the look in one sentence and reuse that sentence in every request. Reference images help on models that accept them. Even then, separate requests drift a little, so check a new asset next to your existing screens before you commit it to the project.
It is risky. Image models can misspell words and warp letterforms, and they do not know your real logo. Leave space in the picture and set text in your own design or in code. If you need your actual mark inside a scene, add it afterwards in your design tool instead of asking the model to draw it.
Yes, the same models run in the toolkit today, so you can test any prompt and any model there now. The MCP server is the way to reach them from the editor, and this page describes it in general terms. For questions about setup or access, write to support and they will point you to what is current.
Want to try a prompt before you connect anything? Open the toolkit and test it on the same models, or ask support how to get the MCP server running in Cursor.
Choose your language
