MCP: generate images and videos straight from Claude, ChatGPT or Cursor

The problem: the assistant thinks, but doesn’t do

For the past two years, many creators have worked the same way. They open an AI assistant to think: the concept, the hook, the script, the image description. The assistant is excellent at that. Then comes production, and everything falls apart: copy the prompt, open a generator in another tab, paste, wait, download, come back to show the assistant the result, start again.

That back-and-forth has a hidden cost far bigger than a few seconds of copy-paste. The assistant loses the thread. It doesn’t see what was generated, it doesn’t know the bottle came out with a cap in the wrong colour, it can’t correct the next prompt accordingly. Thinking and production live in two separate worlds.

MCP brings thinking and production into the same window.

What is the Model Context Protocol?

The Model Context Protocol (MCP) is an open standard, launched in late 2024 by Anthropic and since adopted by most major AI players. It defines a common way to plug tools and data into an AI assistant. It is often compared to a USB-C port: before, each assistant had its own extension system, incompatible with the others; with MCP, the same tool plugs in everywhere.

The two sides of the plug

• The client is the app you chat in: Claude, ChatGPT, Cursor, Claude Code. It displays the conversation and asks your permission before using a tool.
• The server is the service that provides the tools. In our case, the ToolAcces Studio, reachable at a web address. It describes each of its tools: its name, what it does, the information it needs.

When you plug in a server, the assistant reads its list of tools and their descriptions. Then, as the conversation goes, it decides on its own when a tool is useful, prepares the call, submits it to you, and reads the response. You have nothing to learn: you talk normally, the assistant does the rest.

Three building blocks: tools, resources, prompts

The protocol knows three kinds of elements. Tools are actions (generate an image, check a balance). Resources are data the assistant can read. Prompts are pre-written playbooks that you trigger like a command, and that guide the assistant step by step. This third block is often overlooked; yet it makes the difference between “generate some images” and a truly consistent campaign.

What it looks like in practice

Here is a typical exchange with the ToolAcces connector plugged into Claude. Two requests, in plain language: an image, then its animation.

Claude · ToolAcces connector

Make me a macro shot of a two-scoop pistachio and strawberry cone held against a blue sky, full sun, vertical format.

generer_image · nano_banana_2 · 9:16

Image generated with Nano Banana 2. The credits have been deducted from your balance.

Perfect. Now animate it: the gelato melting down the fingers.

generer_video · image_depart · 5 s

5-second video generated from the same image.

What matters isn’t the first image: any generator can make an ice cream cone. It is the second request. The assistant kept the image at hand and passed it to the video model itself, without re-uploading anything, without switching tools. The same sequence would have taken five minutes and four tabs by hand.

What it changes for content creation

The assistant becomes a project manager

An AI assistant is better than most of us at writing an image prompt: it knows the photography vocabulary, it is methodical, it doesn’t get bored at the tenth shot. With MCP, that talent is put straight to work: you describe the intent, it writes the prompts, launches the generations, looks at the results, and adjusts.

Context is no longer lost

The client brief, the brand guidelines, the campaign tone: everything you have already given the assistant in the conversation serves every generation. No need to summarise it in each prompt, and no risk of forgetting a piece.

Production happens in batches

A twelve-visual campaign goes out in a single call. A product list becomes a list of visuals. A four-week content calendar becomes four weeks of images, filed in your gallery.

Packshot

Macro

Top view

Model

What the ToolAcces connector can do

Once plugged in, your assistant has the following tools. You don’t need to remember their names: it picks them itself based on what you ask.

The three guided workflows

Besides the tools, the connector provides three prompts, which show up as commands in your client. Each one carries a method proven on our own campaigns, so the assistant doesn’t improvise.

• campagne-produit — twelve consistent visuals around a product, with the lock phrase, the grid of six shot types and the cost announced before launching. The method is detailed in our feature on consistent campaigns.
• avatar-ugc — a character who keeps the same face and voice from scene to scene, for testimonial-style content. It rests on three locks: an identity block, a voice block copied word for word, and a reference image.
• animer-mes-images — turns visuals you have already generated into short animated shots, one movement per shot, cost announced before launching.

Plugging in the connector, client by client

The server address is the same everywhere. You then sign in with your usual ToolAcces account, in your browser: no API key to generate, copy or store.

https://www.toolacces.com/api/mcp

In Claude (app or claude.ai)

1. Open the connector page

   On toolacces.com/en/mcp, the “Connect in Claude” button opens the add-connector window already filled in. Otherwise: Settings → Connectors → Add custom connector, name it “ToolAcces”, and paste the address above.

2. Sign in

   Claude opens a ToolAcces sign-in window. Once connected, the tools show up in the conversation.

In ChatGPT

1. Turn on developer mode

   Settings → Apps and connectors → Advanced. A paid ChatGPT plan is required.

2. Add the remote server

   Under Connectors, add a remote MCP server named ToolAcces with the address above, then sign in.

In Claude Code

One line in the terminal, from any folder:

claude mcp add --transport http toolacces https://www.toolacces.com/api/mcp

Claude Code opens your browser to sign you in on first use.

In Cursor

Settings → MCP → Add new MCP server, or add this to your ~/.cursor/mcp.json file:

{
  "mcpServers": {
    "toolacces": { "url": "https://www.toolacces.com/api/mcp" }
  }
}

Credits, security, control: who decides what

Plugging a paid tool into an assistant that acts on its own raises a fair question: who holds the wallet? The answer comes down to four safeguards.

1. Your approval before each call. MCP clients ask your permission before using a tool. You see what the assistant is about to do, and with which parameters.
2. The cost announced before spending. The connector is designed so the assistant announces the price before any video (the most expensive generation) and before any campaign, thanks to plan_campagne, which prices without charging.
3. The same credits, at the same rate. No parallel billing: these are the Studio credits of your subscription. You can ask “how many credits do I have left?” at any time.
4. Revocable, time-limited access. Remove the connector in your client and access stops. It also expires on its own, and closes for good if the subscription ends.

Five ways to use it day to day

1. The product campaign in one conversation

Describe your product in one sentence and launch the campagne-produit workflow. The assistant writes the reference description, submits it to you, picks the shots, announces the total cost and waits for your green light. Twelve visuals later, you ask it to present them in publishing order.

2. A client brief turned into boards

Paste a client’s brief into the conversation. Ask for three visual directions, two images each. You walk into the meeting with six proposals instead of a moodboard of references found elsewhere.

3. A site’s visuals while you code it

In Cursor or Claude Code, the assistant writing your landing page can also generate the hero image, the section illustrations and the share thumbnail, in the right format, without leaving the editor. No more grey placeholders while you wait for the designer.

4. The avatar that stays the same

The avatar-ugc workflow builds a character from an identity block approved once, then rolls it out across six scenes: bathroom, car, café terrace. Same face, same voice, different sets.

Reference

Terrace

Metro

Office

5. Batch animation

Already have images? Ask the assistant to animate them with animer-mes-images: it picks one movement per shot, favours gesture shots, and gives you the total cost before launching.

Best practices that avoid nasty surprises

• Approve the product description before any batch. A mistake in the reference phrase is paid for twelve times.
• Give the format in your first request. 9:16 for social, 1:1 for a feed, 16:9 for a website. Cropping afterwards loses part of the image.
• Test at low resolution. Validate the idea at 1K, regenerate in high definition only the shots you keep.
• Ask for the cost before videos. The assistant already does it, but “how much will this cost?” never hurts.
• Let videos run. A video takes one to several minutes: carry on with the conversation, the assistant can check its progress.
• One conversation per campaign. Context stays clean, and you can come back to it for a variation.

The limits, honestly

• You need a compatible client. Claude, ChatGPT (paid plan, developer mode), Claude Code and Cursor work. Not every assistant accepts remote MCP servers yet.
• The assistant doesn’t always see as well as you do. It can describe an image, but a label detail or a crooked letter will jump out at you faster than at it. The final check is yours.
• Quality still depends on prompts. MCP doesn’t make a model better: it makes the method easier to apply. That is why the guided workflows exist.
• Video stays asynchronous. It doesn’t arrive instantly in the conversation; you sometimes need to ask how it’s going.

Frequently asked questions

Go further

Complete guideAI tool stack for freelancers and small agencies

ArticleAI ad campaign: twelve visuals, one product, the lock-phrase method

ArticleHow to create a product ad with AI from a single photo

ArticleAI UGC avatar: the same face and the same voice, scene after scene