Connect the Picasso IA MCP server and Claude can create and edit images without you leaving the conversation. You describe what you need, approve the request, and the picture comes back as a file. Draft a hero image for a post, then keep refining it in the same chat.
Most people already have Claude open while they write a post, plan a launch or outline a deck, and the picture is the one thing they have to leave the chat to make. Connect the Picasso IA MCP server and that changes: you ask for the image in the conversation, you approve the request, and the file comes back to the place where you were already working.
The usual routine is a context switch. You finish a paragraph with Claude, open another tab, retype the idea as an image prompt, wait, download, and then explain the result to Claude all over again. Each hop loses detail. The chat already knows the audience, the tone of the article and the sentence the picture is supposed to support, so it is the best place to write the image brief.
MCP, the Model Context Protocol, is an open standard that lets an assistant call outside tools. Clients that speak it, such as Claude Desktop, Claude Code and Cursor, can connect to the Picasso IA MCP server and use the catalog from the conversation. One thing is worth being clear about from the start: Claude does not draw. It writes the prompt and asks for the tool. The picture itself comes from an image model in the catalog, the same one you could pick yourself in the toolkit.
That split is also why you stay in control. The client asks you to approve a tool call before it runs, so nothing is generated, and nothing is spent, behind your back.
Think of it as a short relay. You describe the image in plain words. Claude turns that into a proper prompt and proposes a request to the Picasso IA tool. You approve it. The request goes out under the key from your account, the chosen model generates, and the result comes back as a file you can download. The credits for that generation are the same credits you would spend in the toolkit, and the amount depends on the model and the settings.
Different jobs call for different kinds of model, and you can tell Claude which one you want or let it suggest.
| What you ask for | How to word it | Kind of model that fits |
|---|---|---|
| A hero image for a post | Say the topic, the mood and the shape: wide, with empty space on one side | Text-to-image, such as flux-2-pro or seedream-5-pro |
| Another take on a draft | Name the image you liked and the one thing to change | An editing model, such as nano-banana-pro or qwen-image-edit-plus |
| Edits to a picture you attached | Describe the change, not the whole scene | An editing model that accepts a reference image |
| A sharper, larger file | Ask for the best version of the one you chose | A model from the super-resolution category |
| A few rough directions | Ask for three different looks of the same subject | A fast text-to-image model, then pick one |
The catalog has 502 models, 212 of them text-to-image, so there is room to try the same brief on more than one engine. The full catalog lists them all.
Here is the path from nothing to a picture in the chat. The first step happens once; the rest is how every request goes.
If you want to try a prompt before wiring anything up, the toolkit runs the same catalog today, and the pricing page explains how credits work.
A hero image for a post. You have just finished an article about slow mornings. Ask Claude for a wide image to sit above the title, a quiet kitchen in early light, empty space on the left for the headline, no people. Claude writes a specific prompt, you approve it, and one image comes back. If the light is right but the kitchen is cluttered, you do not start again. You say remove the clutter on the counter, keep everything else, and an editing model works from the picture you already have.
Iterating on a draft. Say you have a rough illustration for a newsletter and the colors feel off. Attach it, tell Claude to keep the composition and move the palette toward warm, muted tones, and approve the edit. Compare the two files side by side. If you like the new palette but lost a detail, ask for that one detail back. Two or three rounds usually settle a picture like this, and every round is a single request that you approved.
Tip: change one thing per request. When you ask for a new palette, a new background and a new pose in one message, you cannot tell which change helped, and the model has more chances to drift away from the picture you liked.
For posts that go to social feeds, the social media posts page covers sizes and layouts, and AI poster design is a good next read when the image has to carry text.
A few habits make the difference between a first image you can use and a long back and forth.
It is honest to say where the setup does not help. Generation takes time, and it costs credits, so a long chat with many approved calls is not free. Models differ: the same prompt can come back sharp on one engine and muddy on another, and Claude cannot see the picture the way you do, so your eyes are the quality check.
Some jobs belong elsewhere. If you need precise layout, exact brand colors or a logo that must match a file you already own, a design tool is the better place, and the AI image is raw material for it. Anything that has to be literally true, such as a product photo for a listing, should come from a real photograph of the real item. Faces are a line too: use your own likeness or fictional characters, never a real, identifiable person.
Finally, a connection is only as good as the client behind it. The client has to support MCP, the key has to be valid, and you still approve each call. If a step fails, support is the place to ask, and the toolkit is always there for generating by hand.
Not by itself. Claude writes text, so it cannot draw. With the Picasso IA MCP server connected, Claude can ask a tool to generate the picture. It writes the prompt and picks a model, the tool runs the request, and an image model from the catalog produces the file. You approve the request before it runs, so you decide what gets generated.
MCP, the Model Context Protocol, is an open protocol that lets an AI assistant call external tools through one common method. Instead of copying prompts between apps, the assistant can reach a tool directly, ask it to do a job and get the result back in the conversation. Picasso IA runs one of these tools, so a compatible client can generate images, video and more from the catalog.
Any client that supports MCP can connect. Claude Desktop, Claude Code and Cursor are examples of clients that do. Each one has its own place to add a server, so follow your client's connection settings and use the key from your account. If your client does not support MCP yet, the toolkit on the website runs the same catalog and needs no connection at all.
Yes. For the MCP server, access works with a key you create in your account. Treat it like a password: keep it out of shared documents and screenshots, and replace it if you think someone else has seen it. The toolkit in the web app works without a key because you are already signed in to your account there.
No. The client asks you to approve a tool call before it runs, and a generation only happens after you say yes. The call uses the same credits as the toolkit, and the amount depends on the model and the settings. Read the proposed request before you approve a large batch, and look at the pricing page to see how the plans work.
The server uses the same catalog as the toolkit, which holds 502 models, 212 of them text-to-image. That includes engines such as flux-2-pro and seedream-5-pro for new images, plus editing models such as nano-banana-pro and qwen-image-edit-plus for changes. You can name a model in your message or let Claude suggest one and read its choice before approving.
Yes, when the request uses an editing model. Attach the image in the chat, say what should change and what should stay, and approve the call. The edit starts from your picture instead of a blank page. Keep your instruction narrow, one change at a time, so you can tell what helped and go back to the earlier version if a round goes wrong.
It takes some time, and it depends on the model and the settings, so no single figure would be right. Heavier models usually take longer than lighter ones. Cost works the same way: the credits depend on the model and the settings you choose. The pricing page explains the plans, and new accounts get free credits to try things out.
Check before you rely on it. Rules differ by model and by country, and the law on machine-generated images is still settling. Keep your prompts and outputs for your records, review each picture before publishing, and read the terms of the platform you post on, since they change. If the stakes are high, ask a qualified professional rather than a chat.
Use only your own likeness, with the consent you would want for any photo of yourself, or fictional characters. Do not generate real, identifiable people who have not agreed to it, and do not present an AI picture as a photo of a real event. Check the output carefully, because faces and hands are where models still make visible mistakes.
Want to see what a prompt looks like before you connect anything? Open the toolkit, pick a model and run the first one today, or ask support if you get stuck.
Choose your language
