AI coding agents have become surprisingly capable.
Give an agent a task and it can write the code, create files, run commands, interact with APIs, and sometimes complete an entire feature without much human intervention.
But there is still a gap I kept running into while building projects with them.
Images.
Suppose I ask an agent to build a landing page for a new product.
It can create the React components, write the CSS, build the API and put everything together.
But when it needs a hero image or product visual, the workflow usually stops.
I have to open another image generation tool, create the image, download it, put it into the project and continue.
That made me wonder:
What if image generation was simply another tool available to the agent?
Using MCP for image operations
This is where MCP becomes interesting.
Instead of building a custom integration for every AI agent, an MCP server can expose capabilities as tools that an agent can discover and use.
I started building an MCP server specifically for image operations.
The first version exposes tools for generating images, editing and inpainting existing images, removing backgrounds, upscaling images and transforming images.
The goal isn't to create another standalone image generator.
The goal is to make image manipulation part of an agent's workflow.
A different kind of workflow
Consider a task like:
Build a landing page for a new sneaker company. Create the hero visual and product images that the page needs.
With a traditional workflow, the developer might generate those images separately and then give them to the coding agent.
With image tools available to the agent, the workflow can potentially become:
The agent understands the page requirements.
It determines that visual assets are needed.
It calls the image generation tool.
The images are returned to the workflow.
The agent continues building the page using those assets.
That small difference is important.
The agent isn't just generating an image because the user explicitly asked for one.
It can use image generation as part of completing a larger task.
Why I think this matters
We're moving from AI assistants that answer questions toward agents that actually perform tasks.
As agents become more capable, they'll need access to more than just coding tools.
They'll need browsers.
They'll need databases.
They'll need file systems.
They'll need APIs.
And I think they'll also need tools for creating and manipulating visual content.
An agent building a website shouldn't necessarily have to stop because it needs an image.
An agent creating a presentation shouldn't necessarily need a human to create the illustrations.
An agent working on an ecommerce site could potentially create product visuals as part of the workflow.
Image generation becomes less of a separate application and more of a primitive that an agent can use.
Building ImageMCPServer
I built ImageMCPServer around this idea.
It provides image capabilities through MCP and is designed to work with MCP-compatible agents and tools such as Claude, Cursor and other agent environments.
The project is still early, and I'm particularly interested in figuring out which image operations are genuinely useful in agent workflows.
Generation is obvious.
But editing, inpainting, background removal and upscaling may actually become more useful when an agent is working with existing assets.
For example, an agent could take an existing product image, remove its background, generate a new environment around it and upscale the final result without the developer manually moving the file between different applications.
What's next?
The interesting part for me isn't simply adding more image models.
It's figuring out what an agent should actually be able to do with visual content.
Should agents be able to create entire sets of brand assets?
Should they be able to inspect an existing design and modify its images?
Should image tools expose higher-level operations rather than individual generation APIs?
Those are some of the things I'm exploring now.
The project is available at:
If you're building AI agents, I'd be interested to hear what you think.
What image capability would make the biggest difference in your agent workflow?
Disclosure: I used AI assistance while editing this article, but the project, examples, and technical direction described here are based on my own work and experience.
Top comments (0)