Asset Prompter
Srpski SR Star on GitHub Install

Five ways to get generated images and video into your coding agent's project

Claude Code, Codex and Cursor know which pictures a project needs. How the files get there depends on your generator. Here are the five ways, honestly compared.

By Đorđe Novaković, who makes Asset Prompter · Updated

The short answer

If your agent has an image tool built in and its model is good enough, use that; Claude Code has none of its own. If your generator has an official MCP server on your plan, connect it. If it has only an API, you can pay for it per image. If it has none, as with Midjourney and Google Flow, you carry the files yourself: by hand, or through a shared folder like Asset Prompter's. The ways below go from the least of your own work to the most; the facts behind each are in the table of generators and agents, checked on 8 October 2026.

1. The agent's own image tool

Codex has a built-in image_gen tool, Cursor's agent has generated images since version 2.4, Antigravity's agent decides by itself when to call an image model, and Gemini CLI has the nanobanana extension. Claude Code has no image generation of its own: Anthropic's help centre says Claude doesn’t generate photos or illustrations the way image-generation tools do.

Who presses Generate
The agent.
What you pay on top
On Codex, nothing on top of a ChatGPT Plus, Pro or Business plan until its Codex limits run out: image turns use them faster than other turns, and after that image generation draws on credits. Image generation is not on the Free plan. Gemini CLI's extension needs a Gemini API key, so API billing. Cursor and Antigravity don't say.
What the agent sees
The picture it made.
Where the versions live
Where the tool saves them. Cursor saves to the project's assets/ folder by default. Codex keeps them under $CODEX_HOME/generated_images/, and its image skill tells the agent to copy what the project uses into it, as a new file such as hero-v2.png rather than over the old one.
Pick it when
The agent's model is good enough for the picture, and you don't need a particular generator.

2. A generator's MCP server

Higgsfield, Ideogram, Krea, Kling and Runway have MCP servers you sign in to with your account; Leonardo.Ai has one that runs on its API, and Luma's GitHub organisation has one that runs on Luma's.

Who presses Generate
The agent.
What you pay on top
Higgsfield, Ideogram and Runway draw on your plan or its credits, and Krea on your workspace's compute units when you sign in. Leonardo's and Luma's servers need an API key, and use API credits.
What the agent sees
What the server gives back.
Where the versions live
Wherever the server or the agent saves them.
Pick it when
You want batches that run while you are away. As the home page says: for those, pick a generator with an API or an MCP server.

3. Paying for an API

OpenAI's image API, Google's Gemini API (Nano Banana for images, Veo for video), xAI's Grok Imagine, Leonardo.Ai and Luma: the agent writes a script against the API and runs it.

Who presses Generate
The agent, through the script.
What you pay on top
Each generation, billed apart from any app plan you already pay for. For Grok Imagine, xAI's own pages don't say whether a Grok plan covers any of it.
What the agent sees
The files the script saves.
Where the versions live
Where the script saves them; ask for the project folder, with a new name per version.
Pick it when
You want a pipeline you can run again, many sizes at once, or batches. Keep the API key out of the repository.

4. Copy and paste by hand

Any generator: you take the prompt from the chat to the generator, and the file back to the project.

Who presses Generate
You.
What you pay on top
Nothing.
What the agent sees
Nothing, unless you attach the file or describe it.
Where the versions live
Wherever you saved them.
Pick it when
You need one or two pictures, once. Then this is fine, and you need nothing else.

Past a few pictures, it goes the way the home page draws it in “Without it, you are the courier”:

  • You copy a prompt out of the chat, paste it into the generator, download the result, rename it and move it into the repo.
  • The agent never sees what came out, so you describe the picture back to it in words.
  • Three rounds later nobody knows which file was the good one, or which prompt made it.
Pixel art: a person runs between a chat bubble and a pile of pictures, papers flying, while a robot waits and sees none of it.

5. A shared folder: Asset Prompter

A small local app that carries prompts and results between the agent and any generator you use by hand, including Midjourney and Google Flow. It works with any agent that can read and write files in the project folder.

Who presses Generate
You, every time, in your own generator.
What you pay on top
Nothing: the app is free, under the MIT licence. You pay for your generator and your agent, as now.
What the agent sees
Every result. It reviews each one, and reads a clip through its frame sheets and motion map, made with ffmpeg, which setup installs on Windows and Linux.
Where the versions live
In the slot folder, as plain files: every version, result, change request and review. Your approval is final.
Pick it when
Your generator has no API, or you want the last word on every picture.

Setup covers Windows and Linux; on macOS you install Bun yourself. The six steps on the home page show one picture going through it.

A slot's card with your picture in it, a dining room on a space station. The card is marked Agent's turn, with the buttons Approve v1 and Request changes.
A slot's card after you dropped your result on it: now it is the agent's turn.

Side by side

“Depends” means it is up to the tool or to what you ask the agent for.

The five ways compared, from the facts in the table of generators, checked 8 October 2026.
Way Who presses Generate Extra cost Runs while you're away Agent sees the result Versions kept as files Works with Midjourney and Flow
The agent's own tool The agent Depends on the agent Yes Yes Depends No
A generator's MCP server The agent Your plan, or API credits Yes Yes Depends No
An API The agent Each generation Yes Yes Depends No
By hand You Nothing No Only what you show it Only what you keep Yes
Asset Prompter You Nothing No Yes Yes Yes

When you don't need Asset Prompter, and when it helps

It doesn't generate anything. You press Generate in your own tool, every time, so it is not the answer to every case.

You don't need it if

  • Your agent's own image tool, or a generator's MCP server on your plan, does the job, and you are happy to let the agent pick the results.
  • You need batches that run while you are away.
  • You need one or two pictures, once.
  • You work in a chat window in a browser rather than with an agent that works with files: the app needs one that can.
  • You are on macOS and won't install Bun yourself.

A good fit if

  • Your generator has no API, and you want your agent to see the results.
  • You already pay for a generator plan and want no API bills on top.
  • You want the last word. Nothing is done until you approve it.
  • You want it on your machine: plain files in a folder, no account.
  • You want the agent to check your video clips, through frame sheets and a motion map (with ffmpeg, which setup installs).

Give your agent a way to ask for pictures.

Download the folder, run setup, paste one message.