Photoshop MCP Tutorial: Control PS with Natural Language for Retouching and Background Removal

Photoshop MCP lets AI Agents control Photoshop directly via MCP protocol using natural language.
Photoshop MCP is an unofficial open-source project that wraps Photoshop's functionality into 118 tool interfaces via the MCP protocol, allowing AI Agents like Codex to perform layer operations, background removal, and color adjustments through natural language. No Adobe developer account or paid AI service is required, and it supports both Windows and macOS. Setup involves confirming Node.js 18+, installing the MCP service, and adding config to Codex's config.toml. Background removal works well in testing, though complex edge handling remains a limitation — best suited for exploration and lightweight tasks.
Let AI Directly Control Photoshop
As AI Agent capabilities continue to expand, more and more developers are exploring how to let large language models directly control desktop software to get real work done — not just chat. This tutorial introduces an open-source solution that deeply integrates AI Agents with Photoshop: Photoshop MCP.
With it, we can have Codex (or any other AI Agent that supports the MCP protocol) directly call Photoshop's tool interfaces, enabling layer operations, text editing, color adjustments, selections and masks, background removal, and a full range of professional retouching actions. The entire interaction requires nothing more than describing the desired effect in natural language — the AI handles all the analysis and execution automatically.
It's worth noting upfront that Photoshop MCP is an unofficial community project with no affiliation or authorization from Adobe. It currently supports both Windows and macOS, and offers a remarkably broad set of Photoshop capabilities.
Low Barrier to Entry: Key Things to Know
Before getting started, here are a few user-friendly aspects of this project worth highlighting:
- No Adobe Developer account required
- No additional AI service purchases needed
- The project author claims support for all Photoshop versions
While the author claims compatibility with all versions, practical testing suggests using Photoshop 2023 or later for better stability and more complete tool support.
Environment Setup and Installation
The core of the setup process is establishing the "AI Agent → MCP Protocol → Photoshop" pipeline. Before starting, confirm your base runtime environment is ready.
Check Your Node.js Environment
The first step is verifying your local Node.js installation — version 18 or higher is required. If your version is below 18, you'll need to upgrade Node.js first, or the installation will fail.

Once your environment checks out, open Photoshop and create a blank document so you can verify the connection later.
Install the Photoshop MCP Service
Next, run the installation command from the project documentation in PowerShell. Photoshop MCP includes a standalone WebUI, but to align with modern Agent workflows, this tutorial skips the standalone UI and connects directly to Codex.
Once the installation logs start scrolling, the installation has essentially succeeded. The logs show that MCP has registered 118 Tools and 23 Prompts — these are the capabilities exposed to the AI Agent for calling.

Some warning messages will appear during installation. Here's how to interpret them correctly:
- Warn warnings: These do not indicate installation failure. They're notifications about a deprecated dependency package and can be ignored.
- Status 400: Also ignorable. This simply means anonymous telemetry data failed to upload (because Codex configuration isn't complete yet) and has no impact on functionality.
At this point, do not close the PowerShell window.
Configuring Codex Integration
After installation, the real "connection" step is getting Codex to recognize and connect to the MCP service.
Edit the config.toml Configuration File
Locate Codex's user data directory and open config.toml in a text editor. Add the Photoshop MCP configuration under the mcp_servers node. After saving, quit Codex and restart it so it picks up the new configuration.

After restarting, open Codex's settings and navigate to the plugins section to view the MCP list. If you can see the Photoshop MCP entry you just added, the configuration has been recognized correctly. At this point, you can close the PowerShell window — no need to manually open any terminal going forward.
Verify the MCP Connection
Don't rush into operations after configuration. First, have Codex check whether it can access MCP and list the available Photoshop tools. A few authorization prompts will appear — allow all of them.
From the output, you should see:
- Codex successfully accessed MCP
- The local Photoshop version information was detected
- UXP Bridge status shows as reachable
This confirms that the plugin service has started successfully and is listening on the specified port — the MCP connection is officially established. On the Codex side, all 118 callable tools are listed. To clarify a common misconception: these are not 118 separate software features, but rather the tool interfaces MCP exposes to the AI Agent — essentially the "controls" the AI uses to understand and operate Photoshop.
Hands-On Demo: Natural Language Retouching
With the connection established, it's time for the most interesting part: putting it to work.
Reading Document Info and Creating Layers
First, have Codex connect to Photoshop MCP and retrieve the current Photoshop state and open document information. It successfully reads the document name, dimensions, and other details.

One thing worth noting: the document name appears as garbled text in the output. The likely cause is an encoding compatibility issue, and no fix is currently available — we'll need to wait for the project author to address it. This is a reminder that community projects still have room to grow in terms of polish.
Next, a quick test asks Codex to create a new layer named "AI Test." The layer is created without issue, confirming that write operations work as well.
AI Background Removal via Natural Language
The most compelling test is background removal. Take a cartoon character image with a solid blue background — the goal is to remove the blue background and keep the character. Import the image into Photoshop, then send a natural language instruction.
After a short wait, the AI successfully removes the blue background with the character largely intact. From image analysis to actual execution, the entire workflow is handled automatically by the AI — the user simply describes the desired outcome in plain language.
That said, it's important to be realistic about its limitations: the current automation cannot achieve a high level of precision, and may leave artifacts when dealing with complex edges, hair strands, and other fine details.
Summary and Takeaways
Photoshop MCP demonstrates the enormous potential of the MCP protocol in the direction of "AI controlling professional software." It wraps Photoshop — a highly specialized, complex-to-interact-with application — into standardized tool interfaces for AI Agents to call, turning "talking your way through photo editing" from a concept into a usable reality.
From an engineering standpoint, the barrier to setting this up is quite low — no developer account needed, no paid AI service required, dual-platform support, and deployable by any user with basic technical ability. From a maturity standpoint, however, issues like garbled encoding and imprecise edge handling signal that it's still in early stages. It's better suited as a tool for technical exploration and lightweight tasks than as a replacement for professional production workflows.
For developers focused on AI Agent deployment in the real world, the biggest takeaway from projects like this is: MCP is becoming a universal bridge connecting AI with the desktop software ecosystem. We can expect more and more professional applications to integrate AI capabilities through similar approaches in the future.
Related articles

Insufficient Source Material to Generate a Valid Article
The provided source material is a single unrelated tweet with no AI or tech relevance — insufficient to support a complete, valid technical article.

Insufficient Source Material to Generate a Valid AI/Tech Article
This source material is a tweet about the ages of Underworld members — unrelated to AI or tech, and insufficient to support a full article.

Insufficient Material: Unable to Generate a Valid AI/Tech Article
The provided material is a condolence tweet about a San Diego mosque attack — unrelated to AI/tech and too limited to generate a valid technical article.