Expose image, video and voice generation as tools your agent can call. The agent decides what to make, sends the request, and checks the file that comes back before it carries on. You set the approval steps and the limits, and you can test every prompt in the toolkit first.
An agent that only writes text hits a wall the moment a task needs a picture, a clip or a voice. Media tools remove the wall. The agent decides what to make, calls a tool with a prompt, gets a file back and looks at it before moving on. The Picasso IA MCP server is built around that pattern: image, video and voice from the same catalog the toolkit runs, available to any client that speaks MCP.
MCP, the Model Context Protocol, is an open protocol that lets an AI assistant call external tools. A client such as Claude Desktop, Claude Code or Cursor connects to a server, sees the tools it offers and uses them in the middle of a conversation. With the Picasso IA server, those tools generate media instead of fetching data.
The practical gain is that the agent stops handing you a to-do list. Ask it for a product announcement and it can draft the copy, make the header image, cut a short clip and record the voiceover in one run, then show you all of it together. You review a finished set instead of copying prompts between five tabs.
Behind the tools sits the same catalog as the toolkit: 502 models, including 212 for text-to-image, 121 for text-to-video and 24 for text-to-speech. The same credit system applies, so every generation costs credits that depend on the model and the settings. Access works with a key you create in your account, and you should treat that key like a password. The catalog shows what the agent can choose from, and the pricing page explains how credits and plans work.
An agent that uses media tools well follows the same three beats every time. It decides what to make, calls the tool, and checks the result. The third beat is the one people skip, and it is the one that keeps bad output from reaching a user.
Deciding is where your instructions matter. An agent told to illustrate this article will invent its own idea of what that means. An agent told to make one 16:9 header image in a flat illustration style, with no text in the picture, has a brief it can follow and a standard it can be held to. Write the brief the way you would for a freelancer who cannot ask questions.
Calling is one request per generation. The agent sends a prompt and any attached files, waits while the model works, and receives the result as a file you can download. Generation takes time, longer for video than for a still image, so a well-built agent waits for one result before it plans the next step rather than queueing ten and hoping.
Checking is where the agent compares what came back with the brief. It can look at a still image, read the transcript of a voice clip, or sample a few frames of a video, and then decide to accept, retry with a changed prompt, or stop and ask you.
Image, video and voice are not interchangeable, and neither are the models inside each group. The table below pairs each medium with the decision the agent makes and the check that should happen before it moves on. The model names are starting points you can try in the toolkit, not a fixed recommendation.
| Medium | What the agent decides | Check before moving on | Models to try |
|---|---|---|---|
| Image | Subject, style, aspect ratio | No stray text, correct count of objects, hands and faces intact | flux-2-pro, nano-banana |
| Image edit | Which element to change and which to keep | The untouched parts really are untouched | nano-banana, flux-kontext-pro |
| Video | Shot length, motion, camera move | Motion looks natural, the subject stays consistent across frames | veo-3.1-fast, kling-v3-video |
| Voice | Script, voice, pacing | Names and numbers pronounced correctly, no clipped words | speech-2.8-turbo, qwen3-tts |
| Music | Mood, length, instrumentation | Fits the video it will sit under, no abrupt ending | lyria-3, music-1.5 |
Two habits make the table work in practice. First, let the agent pick the cheaper, faster model for drafts and a stronger one only for the final version, and say so in its instructions. Second, keep a short list of models the agent is allowed to use. A wide-open catalog is a lot of choice for a system that cannot taste the difference, and a list of three or four makes its results more consistent from one run to the next.
This is the journey from an idea to an agent that makes media on request. It assumes you already have an MCP-capable client and an account.
The server gives the agent tools. It does not decide how much freedom the agent has. That part is yours to configure in your agent or client, and it is worth doing before the first real task, because an agent that can spend credits without asking will eventually spend them on a loop you did not foresee.
Treat the credit balance as the agent's budget. Tell it the maximum number of generations per task, tell it to stop and ask when a request would go past that number, and keep video generations behind a manual approval, because they cost the most credits and take the longest.
Start strict. A sensible first setup asks for approval before every video, allows a few image attempts per task, and requires the agent to show you the result before it uses a file anywhere public. Many clients can ask you to confirm each tool call, which is the simplest approval step there is.
Loosen it one rule at a time. If images are reliably good after fifty tasks, let the agent run them without asking. Keep a human in front of anything with a face, a voice or a claim attached to it. Use only your own likeness or fictional characters, never a real, identifiable person, and never clone a voice you do not have the right to use.
An honest page says where the tool stops being useful. For agents that generate media, there are five places:
None of this is a reason to skip the tools. It is the reason the agent drafts and a person signs off.
It means the agent can call a tool that generates an image, a video clip or a voice recording, the same way it might call a search tool. The agent decides what to make, sends a request with a prompt, and gets a file back to inspect. Without the tools, it can only describe the media in words and leave the rest to you.
MCP, the Model Context Protocol, is an open protocol that lets an AI assistant call external tools through a standard connection. Because it is a standard, one server works with any client that supports it, such as Claude Desktop, Claude Code or Cursor. The Picasso IA MCP server uses it to put image, video and voice generation inside the conversation.
The agent works with the same catalog as the toolkit, which currently holds 502 models. That covers 212 text-to-image models, 121 text-to-video models, 24 text-to-speech models and 10 for music, among others. You can narrow the list in your agent instructions, which is worth doing so the results stay consistent from one task to the next.
Yes. Access for the MCP server works with a key you create in your account, and the key acts on your behalf and spends your credits. Treat it like a password: keep it out of shared chats, public repositories and screenshots, and replace it if you think it has leaked. The toolkit itself needs no key, because it is a workspace in the web app.
It follows your instructions. If you say that blog posts get a header image and that product launches get a short clip and a voiceover, it will pick the medium from the task. The clearer the rule, the fewer surprises, so write down which medium belongs to which job and what the agent should do when a task fits none of them.
Every generation costs credits, and the amount depends on the model and the settings the agent chooses. Video usually costs more than a still image, and a stronger model costs more than a fast one. This page does not quote numbers because they change, so check the pricing page for plans and use the toolkit to see what a specific model costs before you put it in your agent.
Put the limit in its instructions and use your client's approval prompts. Set a maximum number of generations per task, require confirmation before any video, and tell the agent to stop and ask when a request would go over the cap. Start with tight rules, then relax them one at a time as the results earn your trust.
Partly. It can inspect a still image, read a transcript of a voice clip and compare the output with your brief, and that catches obvious failures such as stray text or a missing object. It cannot judge taste, brand fit or whether a voice sounds right for your audience, so keep a person in the final review for anything public.
Only with your own likeness, or with fictional characters. Never generate a real, identifiable person, and never clone a voice you do not have the right to use. If the agent might receive photos or audio from other people, say in its instructions that it must refuse those requests and ask you instead.
Open the toolkit and run the same prompts there. It uses the same catalog, so you can compare models, see what an image or a clip looks like, and listen to a voice sample before you write a single instruction. Questions about setup or your account go to support.
Want to see what your agent would get back? Open the toolkit, run the prompts you plan to give it, and judge the results first.
Choose your language
