Connect your AI assistant to the Picasso IA catalog and ask for an image in the same conversation where you are writing, planning or coding. Claude Desktop, Claude Code and Cursor call the models for you, and the result comes back as a file you can download.
Ask an AI assistant to write your landing page and it will, but ask it for the hero image and you get a placeholder or an apology. MCP closes that gap. It is an open protocol that lets an assistant call outside tools, and the Picasso IA MCP server is one of those tools: a connection that lets Claude Desktop, Claude Code and Cursor create images from the catalog without you leaving the conversation.
Most image work is not one big creative session. It is a dozen small moments inside other work: you are drafting a blog post and need a cover, you are building a prototype and need placeholder art, you are planning a campaign in a chat and want to see the idea before you commit to it. Each time, the usual routine is the same. Copy the prompt, open another tab, paste, wait, download, go back and explain to the assistant what you got.
An MCP connection removes the copying. The assistant that already knows your project, your tone and your constraints writes the prompt itself, asks the model for the image and hands the result back in the same thread. You keep the context, and the assistant keeps it too, so a follow-up like make the second one warmer and less cluttered actually means something.
The catalog behind it is the same one that runs in the toolkit: 502 models, 212 of them text-to-image, with the rest covering video, voice, music, 3D and editing. What changes is where you sit. In the toolkit you drive the models yourself. With the MCP server, your assistant drives them for you.
MCP stands for Model Context Protocol. Think of it as a shared plug standard. An assistant that speaks MCP can be connected to any server that speaks it too, and the server tells the assistant what it can do. A calendar server might say it can read events. A files server might say it can list folders. The Picasso IA server says it can generate media from the catalog.
Two things follow from that. First, the assistant decides when to use the tool, based on what you ask. If you say I need a cover image for this article, it can reach for the image tool without you naming a model. Second, you stay in charge of the account side: the connection uses an API key from your account, each generation is one request, and the credits come out of your balance the same way they do in the toolkit.
None of this requires you to understand the protocol in depth. If you can edit a settings file or paste a key where a client asks for one, you can connect it. If you cannot, the support page is the place to ask.
The same connection serves very different jobs depending on where you are working. This table compares the common places people reach for it, what you would ask, and what to expect back.
| Where you work | What you ask for | What comes back | Best for |
|---|---|---|---|
| Claude Desktop | An image while you chat through an idea | A file you can download | Brainstorming visuals and moodboards in conversation |
| Claude Code | Placeholder art or a hero image while building | A file you move into your project | Prototypes, docs and landing pages in progress |
| Cursor | An icon, banner or illustration for the app you are editing | A file you drop into your assets folder | Filling a design gap without switching apps |
| Another MCP-capable client | The same request in the same plain words | The same kind of file | Trying the connection where you already work |
| The toolkit, no assistant | A model you pick and compare by hand | Results in your workspace | Choosing a model before you let the assistant choose |
Notice the last row. The toolkit is not a lesser version of the MCP route, it is the testing bench for it. If you are unsure which model suits your style, run a few prompts there first, then tell your assistant which one to prefer.
The exact screens depend on the client you use, so this walkthrough stays at the level of what you do, not which button to press. None of it needs more than a few minutes of attention.
You are now prompting twice: you brief the assistant, and the assistant briefs the model. The best results come from giving the assistant the purpose, not just the subject. A cover for a finance newsletter and a cover for a kids birthday invitation might share the subject balloons and still need opposite images.
Give four things: what the image is for, the mood, the format (wide, square, tall) and what to avoid. The assistant turns those into a model-ready prompt and can show you the prompt it sent, which is worth reading. When an image misses, you can see whether your brief was vague or the model ignored a clear one, and that tells you whether to rewrite the request or switch models.
Pick the model by job. FLUX 2 Pro and Seedream 5 Pro are strong general choices, Ideogram v3 Quality is the one to try when the picture contains lettering, Imagen 4 suits clean photographic scenes, and nano banana and Qwen Image Edit are the ones to reach for when you are changing an image you already have instead of starting fresh.
Tip: ask the assistant to produce two prompts with different moods before it generates anything. Reading two short prompts costs nothing, while every generation spends credits, so you only pay for the direction you picked.
For subjects that overlap with other use cases, borrow their habits: the briefing style from AI social media posts works for cover images, and AI poster design shows how to ask for layouts that leave room for text.
A connected assistant is convenient, not magical. These are the places the route falls short, and knowing them saves credits.
Faces and likenesses deserve their own line. Use your own face, or fictional characters. Do not ask an assistant for images of a real, identifiable person who has not agreed to it.
It is a small service that follows the Model Context Protocol and offers image generation as a tool an AI assistant can call. Instead of you copying a prompt into another app, the assistant sends the request to the Picasso IA catalog, waits for the model, and returns the finished image as a file in the conversation. The protocol is open, so any client that supports it can use the same kind of connection.
Any MCP-capable client is the target. The ones this page is written around are Claude Desktop, Claude Code and Cursor, because those are the places people already work while writing, planning and building. Support differs from client to client and changes over time, so check your client's documentation for how it registers MCP servers, and ask the support page if something does not connect.
No programming is needed, but you should be comfortable with a settings screen or a small configuration file, and with pasting an API key from your account where the client asks for it. Claude Code and Cursor users usually find this familiar. If you have never connected a tool to an assistant, expect the first attempt to take a little longer, and reach out to support if you get stuck.
Yes. The server uses the same models and the same credit system as the toolkit, so each generation costs credits that depend on the model and the settings you choose. New accounts get free credits to start. Current plans live on the pricing page, which is the only place worth trusting for numbers. A good habit is to tell the assistant how many images you want before it starts.
The same catalog the toolkit runs: 502 models, with 212 in text-to-image alone. For images, that includes FLUX 2 Pro, Seedream 5 Pro, nano banana, Imagen 4, Ideogram v3 Quality and Qwen Image Edit. You can name one in your request, or leave the choice to the assistant. You can browse everything on the models page before you decide.
The catalog covers those categories, with 121 text-to-video models, 24 text-to-speech and 10 music models, and an assistant connected to it can ask for them in the same way. Expect video to take longer and cost more than a still image, because it generally does in the toolkit too. Start with an image, check the result, and only then move to motion.
Where you sit. In the toolkit you pick the model, write the prompt and compare results yourself. With the MCP server your assistant does those steps from inside your chat or editor, and it already knows what you are working on. The toolkit is better for comparing models side by side, and the MCP route is better when the image is a small step inside bigger work.
Yes, and it is a good idea. The toolkit runs the same catalog today, so you can try any prompt or model there, see what you get and learn which model suits your style. Then you can tell your assistant exactly which one to use. Open the toolkit, run a few prompts, and bring the winners into the conversation as examples.
Treat the key like a password. Put it only in the place your client asks for it, never in a file you commit to a public repository, and never paste it into a chat as plain text. Anyone holding it can spend your credits. If you think it leaked, contact support right away so the account can be checked and the key replaced.
Only of yourself, or of fictional characters. Do not ask for images of a real, identifiable person who has not agreed to it, and do not use generated pictures to suggest someone endorsed or did something they did not. The assistant passes your request to the model, so the responsibility for what you ask stays with you. If you are unsure whether a use is acceptable, ask support.
Ready to try the models your assistant will be calling? Open the toolkit and run a first prompt now, or write to support if you have a question about connecting.
Choose your language
