Ask for an app icon, a landing page hero or a set of placeholder images while you are coding, and get the files back inside the same Claude Code session. The Picasso IA MCP server runs image models from the catalog and you review every result before it ships.
Every project reaches the moment where the app needs an icon, the landing page needs a hero image and the empty states need something better than grey boxes. You can generate all of it without leaving the terminal. The Picasso IA MCP server lets Claude Code call image models from the same session where you are writing code, and the results come back as files you can save into the project folder. You still decide what ships.
Most developers handle visual assets badly, and not because they lack taste. The usual routine is to stop coding, open a design tool or a stock site, search, download, rename the file, move it into the repo and then remember what the original prompt was. Each of those steps is small. Together they are why a prototype keeps its grey placeholder boxes for three weeks.
Keeping the request inside the coding session removes the handoffs. Claude Code already knows what your project is: the framework, the folder layout, the component that is waiting for an image, the colors in your stylesheet. When it asks an image model for a hero picture, it can write the prompt from that context instead of making you re-explain the product in a separate browser tab. The file lands where the component expects it, and the prompt that produced it can sit in a commit message or a comment.
The second reason is range. Picasso IA runs 502 models behind one interface, 212 of them text-to-image, so an icon, a photographic hero and a flat illustration can each go to the model that handles that look best, without you holding a separate account for every engine. The full catalog shows what is available, and the same models run in the toolkit if you would rather test a prompt in the browser first.
MCP, the Model Context Protocol, is an open protocol that lets an AI assistant call external tools. The Picasso IA MCP server is one of those tools: once an MCP-capable client such as Claude Code is connected to it, the assistant can generate images, video, voice and more from the catalog during the conversation.
Setup is deliberately short. You create an API key in your account, add the Picasso IA server to Claude Code following the client's own instructions for adding an MCP server, and give it the key. Treat that key like a password. Keep it out of the repository, out of screenshots and out of shared shell history, and store it the way you store any other secret for the project. If it leaks, replace it from your account.
Generations use the same credit system as the toolkit, so each image costs credits that depend on the model and the settings you choose. New accounts get free credits, and the pricing page is the only place that lists plans. Nothing here needs a spec sheet: one request produces one generation, and the result comes back as a file you can download.
Not every project asset is a good fit, and not every model suits every asset. This table pairs the common requests with models in the catalog that are worth trying first, plus the check that matters before the file goes into a commit.
| Asset | What to ask for | Models worth trying | Check before shipping |
|---|---|---|---|
| App icon | One centered symbol, flat shapes, plain background, no text | recraft-v4-svg, flux-2-pro | Legibility at small sizes, resemblance to existing icons |
| Landing page hero | Wide scene with empty space where the headline will sit | seedream-5-pro, imagen-4, flux-2-pro | Brand fit, odd details, crop on narrow screens |
| Placeholder art | Simple, neutral images that match the layout dimensions | p-image, nano-banana | Replace or review before launch |
| Transparent sprite | Subject on a plain background, then background removal | nano-banana, bria/remove-background | Clean edges on hair, glass and thin lines |
| Soft or small result | Upscaled copy of an image you already like | p-image-upscale | New artifacts introduced by the upscale |
Two habits save credits. First, generate the icon at a few different looks before committing to one, because a style that reads well at 512 pixels can turn to mush at 32. Second, write down the dimensions of the slot before you ask. A hero image for a wide banner and one for a narrow card are different pictures, and asking for the right aspect ratio is cheaper than cropping a wrong one.
For deeper looks at specific asset types, the pages on vector graphics and SVG icons, transparent PNGs and game assets go into prompts and model choices in more detail.
This is the whole loop, from a blank folder to files you would be willing to commit.
A prompt for a project asset is closer to a short design brief than to a poem. Say what the image is for, what must be in it, and what must not. Claude Code can fill in the surrounding context from your project, but you are the one who knows whether the product is playful or serious, so say so in the prompt.
For icons, ask for one subject, flat or lightly shaded shapes, a plain background and no text. Models still struggle with letters inside images, so leave wording to your actual interface. For hero images, describe a scene and leave room: a wide composition with a calm area on the left where the headline will sit gives your layout something to work with. For placeholders, keep the subject neutral and consistent so that the page looks coherent while real photography is still being arranged.
Generate three variants before you decide, and judge them at the real display size. An image that looks great full screen in a file preview can be unreadable as a 32 pixel favicon or too busy behind a line of white text.
When a result is close but wrong, change one thing per round. If you rewrite the whole prompt every time you lose track of what helped. And if you need a set that looks related, such as five feature illustrations, reuse the same style sentence in every request and keep the subject as the only variable. The effects page is a quick way to find style words that models respond to.
This workflow is good at getting the first usable version of an asset. It is not a replacement for a designer, and it has limits you should plan around.
None of that is a reason to skip the tool. It is the reason the review step in the five-step loop is not optional, and why this fits placeholders, prototypes and first drafts better than final brand work.
Claude Code works with text and code, so for pictures it needs an image model behind a tool. The Picasso IA MCP server provides that: once connected, Claude Code can call image models from the catalog during the session. You describe the asset, it makes the request, and the result comes back as a file you can save into your project and review before using it.
An account, an API key created in that account and the MCP server added to Claude Code through the client's own settings for MCP servers. Treat the key like a password and keep it out of your repository. Generations are paid in credits, and new accounts start with free credits. If something does not connect, support can help with the specifics of your setup.
Vector-oriented models such as recraft-v4-svg suit flat symbols, and general models such as flux-2-pro or nano-banana can also produce clean icons when you ask for one centered subject on a plain background. Try two or three, then check each at its smallest real size. Pick the one that stays readable, not the one that looks best in a large preview.
Yes, and wide compositions are a natural fit. Describe the scene, state the aspect ratio of your banner and ask for empty space where the headline will sit. Models such as seedream-5-pro, imagen-4 and flux-2-pro are worth comparing. Review the crop on narrow screens too, because a hero that works on a desktop can lose its subject on a phone.
Each generation comes back as a file. In practice you tell Claude Code which folder in the project should hold it, and the file is saved there so a component can reference it. If you want reproducibility, put the prompt and the model name in your commit message or in a short note next to the asset.
Each generation costs credits, and the amount depends on the model and the settings you pick, so there is no single figure to quote. New accounts get free credits to try the workflow. Current plans and what they include are on the pricing page, which is the only place worth trusting for numbers.
It can be, but it is your call and it needs checking. Review the terms that apply to the model and to your account, look at the rules of the app store or platform you publish on, and confirm that nothing in the image resembles an existing brand or a real person. For anything legal or contractual, ask a qualified professional.
It can produce logo concepts, which is useful for exploring directions. It is weak at exact lettering and precise geometry, so a final logo usually needs redrawing in a vector editor. For brand-critical marks, use the output as a sketch for a designer. If you only need a symbol for a prototype or an internal tool, a generated icon is often enough.
Those tools live in the same catalog, including background removal and upscaling models, so you can ask Claude Code to run an image through them after generating it. Check the result on a dark and a light background, because edges around hair, glass and thin lines are where removal fails. If you want to try the models first, the toolkit runs the same catalog in the browser.
Yes. The toolkit runs the same catalog today: pick a model, write the prompt, attach reference files if you have them and generate. It is the easiest way to learn which model understands your style words, and which settings change the cost. Once a prompt works there, you can ask Claude Code for the same thing inside your project.
Your next hero image can come from the same terminal as your commits. Try a prompt in the toolkit first, or ask support if you want help planning the setup.
Choose your language
