Ask your AI assistant for a video clip and get the finished file back without leaving the conversation. The Picasso IA MCP server connects clients such as Claude Desktop, Claude Code and Cursor to the same video models the toolkit runs, with the same credits behind every generation.
Most AI video work starts with a browser tab, a prompt box and a lot of waiting. An MCP server for video generation moves that loop into the place where you already think out loud: your AI assistant. You describe the clip in the conversation, the assistant sends the request to Picasso IA, and the finished video comes back as a file you can download, made by the same catalog of models the toolkit runs.
MCP, short for 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. MCP-capable clients such as Claude Desktop, Claude Code and Cursor can use it to generate video, and also images and voice, without you copying anything between windows.
The practical gain is context. Your assistant already has the script you pasted, the product description in your repository or the storyboard you drafted ten messages ago. Instead of retyping that into a prompt box, you say make a clip of scene three, and the assistant writes the video prompt from material it already holds. When the clip arrives you react to it in the same thread, and the next attempt starts from what you just saw.
Nothing about the models changes. The server uses the same catalog and the same credit system as the toolkit, where 121 of the 502 models are text-to-video. A generation costs credits, and how many depends on the model and the settings you pick, so the pricing page is where the plans live. Access works with a key you create in your account, and you should treat that key like a password.
Video models are not interchangeable, and part of the appeal of a catalog is choosing the right one per job. You can name the model in your request, or ask the assistant to propose one and explain why. These are three that cover most first projects, all browsable in the catalog:
| Clip you need | Model to try | What to ask the assistant for |
|---|---|---|
| One cinematic scene from a text description | google/veo-3.1 | A single shot with subject, camera move, lighting and sound described in one paragraph |
| A short sequence with several shots | kwaivgi/kling-v3-video | The scene broken into shots, each with its own prompt and length |
| A clip where people speak or sound matters | bytedance/seedance-2.5 | Dialogue written in quotes, plus the ambient sound you want |
| A still photo that starts moving | google/veo-3.1 | The photo as the first frame, with a sentence about the motion |
| The same character across several clips | bytedance/seedance-2.5 | Reference images of the character and a reminder to reuse them in every request |
Models differ in clip length, resolution, aspect ratios and whether they generate audio, so ask your assistant to check the model page before it commits. A clip that works for a vertical social post may be the wrong shape for a landscape presentation, and it is cheaper to discover that before generating than after.
This is the whole journey in plain words, from an empty account to a clip you can review. The setup happens once; after that it is just conversation.
Before step 3, it helps to run your first prompt once in the toolkit. The toolkit runs the same catalog today, so you can see what a model does with your wording and settle on a style before you ask an assistant to repeat it.
The assistant is a good prompt writer when you give it constraints, and a mediocre one when you give it a topic. Compare a clip of a coffee shop with a handheld shot of a barista pouring milk into a ceramic cup, morning window light, steam rising, no music. The second version leaves the model far fewer decisions to guess.
Test every new idea on one short clip before you ask for a full sequence. A wrong direction costs one generation, while a wrong direction discovered after six shots costs six.
A useful brief for any video model covers the same few things: what is on screen, how the camera moves, what the light feels like, how long it should run and what sound you want. If you already have a still image of the subject, offer it as a starting frame, because the model then animates something you chose instead of inventing one.
For longer pieces, ask the assistant to plan first. A list of shots with one line each is easy to correct in text, and it turns a vague film idea into a set of small requests. The same habit powers AI storyboards, where the plan is the product, and it pairs well with AI B-roll when the clips are meant to cut into footage you already have.
Video generation is slow compared with text, and an honest workflow starts from that. Each request is one generation, it runs as a job, and the output arrives as a file you download. Plan around the wait: queue the clip, go back to the script, and look at the result when it lands.
Credits are the other constant. Every generation costs credits that depend on the model and the settings, so a longer clip, a higher resolution or a model with audio can cost more than a short silent test. Ask the assistant to tell you which model and settings it is about to use before it runs something large, and keep a simple habit: draft short, finish long.
Files are the last piece. Because results come back as files, they drop straight into whatever you do next, whether that is a video editor, a shared folder or a project. For a clip that needs polish after generation, video upscaling and turning photos into videos show what the catalog offers beyond the first render.
The server is a way to reach the catalog from an assistant. It is not a different engine, and it does not do more than the models behind it. Five limits are worth knowing before you build a routine on it:
None of this makes the idea weaker. It is the reason the page describes a handoff, not a replacement for an editor: the assistant drafts and requests, the models render, and you decide what ships.
It is a connector that lets an AI assistant call Picasso IA to generate video. MCP is an open protocol for giving assistants access to external tools, and the Picasso IA server exposes the catalog's video models through it. You ask for a clip in the conversation, the request runs as a generation, and the result comes back as a file you can download.
Any MCP-capable client. Claude Desktop, Claude Code and Cursor are examples of clients that speak the protocol. If your assistant supports MCP servers, you add Picasso IA the way you add any other server, using a key from your account. If you are unsure whether a client qualifies, ask on the support page before you build a routine around it.
The same catalog the toolkit uses, which has 121 text-to-video models among its 502. Examples include google/veo-3.1 for single cinematic scenes, kwaivgi/kling-v3-video for multi-shot sequences and bytedance/seedance-2.5 for clips with dialogue and sound. Each model has its own length, resolution and audio options, so check its page in the catalog.
You create a key in your Picasso IA account and give it to the MCP server when you add it to your client. Treat the key like a password: do not paste it into shared chats, screenshots or public repositories, and replace it if you think it leaked. The toolkit and the web app need no key, since you simply sign in.
It costs credits, the same credits as the toolkit. The amount depends on the model and the settings you choose, such as length, resolution and audio, so there is no single price to quote here. New accounts get free credits to start. Current plans are on the pricing page, which is the only place worth trusting for numbers.
Longer than an image, and the time depends on the model, the clip length and the settings, so nobody can give you an honest fixed number. Plan for a job that finishes later rather than an instant reply. While it runs you can keep working in the conversation, and when it finishes the clip is waiting as a file.
The result comes back as a file you can download, like any other output from the catalog. From there it goes wherever your project needs it: a video editor, a shared folder or a draft you send for feedback. Watch every clip before you share it, because the assistant can confirm a file exists but not that the motion looks right.
Yes, with models that support it. Several video models accept a photo as the first frame, and some accept reference images to keep a face, outfit or product steady across shots. Check the model page for what each one allows, and tell the assistant which reference to reuse in every request so the character does not drift between clips.
Only your own, or a person who has clearly agreed, and otherwise stay with fictional characters. Do not generate clips that show a real identifiable person doing or saying something they never did. If a clip will be published, check the current rules of the platform where it will appear, since those change and differ from site to site.
Yes, and it is a sensible first move. The toolkit runs the same models today: pick one, write a prompt or attach a photo, and generate. A few test clips there show you what each model does with your wording, so the prompts your assistant writes later start from something that already worked.
Ready to see what the models do with your idea? Open the Picasso IA toolkit and generate your first clip while the rest of the workflow takes shape.
Choose your language
