Photoshop MCP Server, published by loonghao, connects AI assistants to Adobe Photoshop through photoshop-python-api and the Model Context Protocol. The loonghao/photoshop-python-api-mcp-server repository presents it as a bridge for creating, opening, and saving Photoshop documents, manipulating layers, reading session state, and applying effects or adjustments. In the GitHub check on August 26, 2026, the repository was MIT licensed, had 299 stars and 27 forks, and pyproject.toml listed photoshop-mcp-server 0.1.11 with Python >=3.10,<3.15.
For Educasium, Photoshop MCP should be treated as local automation, not image generation. Its most important constraint is explicit in the README: Windows only. The server depends on Windows COM interfaces to communicate with Photoshop, and macOS or Linux are not supported.
Contents
- What Photoshop MCP does
- Windows, Photoshop, and Python requirements
- MCP configuration and client lifecycle
- Available resources and tools
- Security, files, and limits
- Comparison with ComfyUI MCP
- Educasium position
What Photoshop MCP Does
Key point: Photoshop MCP automates locally installed Photoshop on Windows through COM, instead of manipulating an image file in isolation. The assistant acts inside a real graphics application, with documents and layers.
Main functions
The README says the server can create, open, and save Photoshop documents, create and manipulate text or solid-color layers, retrieve Photoshop session information, inspect active documents and selection, and apply effects or adjustments.
That surface fits repetitive tasks: prepare a document, add text, inspect layer count, create a simple composition, apply an adjustment, or save a version. It is less suited to pure creative generation, which belongs more naturally to ComfyUI MCP or a generative image tool.
What it implies
The server drives Photoshop, so application state matters. Active document, selected layer, Photoshop version, file permissions, and save path can change the result. An assistant claiming it created a layer should be checked in Photoshop or through exposed resources.
The Educasium reflex is to verify session state before and after each action. In Adobe tools, automation can be technically correct and still produce weak visual work.
Windows, Photoshop, And Python Requirements
Key point: the README clearly says the project works only on Windows because of COM interfaces. Photoshop must be installed locally.
System and software
Requirements list Windows OS only, locally installed Adobe Photoshop, and Python 3.10 or newer. The README mentions testing with Photoshop CC2017 through 2024. pyproject.toml specifies Python >=3.10,<3.15, mcp ^1.6.0, photoshop-python-api ^0.24.0, loguru, and tenacity.
This excludes many agency or school environments that run macOS. For those machines, do not promise a simple workaround. The target context is a Windows workstation with Photoshop installed.
Photoshop version
PS_VERSION tells the server which Photoshop version to connect to, such as 2024, 2023, or 2022. That is a practical detail: if several versions coexist or the declared version does not match the installation, the bridge can fail.
A serious integration records Windows version, Python version, Photoshop version, package version, MCP client, and configuration file used. That journal keeps model errors separate from environment errors.
MCP Configuration And Client Lifecycle
Key point: the recommended configuration uses uvx with photoshop-mcp-server and PS_VERSION. The lifecycle depends on local Photoshop and the MCP client.
Standard configuration
The README provides an MCP configuration with command uvx, args --python 3.10 photoshop-mcp-server, and env PS_VERSION. It says this works with Claude Desktop, Windsurf, Cline, and other MCP hosts.
Installation can use pip install photoshop-mcp-server or uv. The console script in pyproject.toml is photoshop-mcp-server, pointing to photoshop_mcp_server.server:main.
Application lifecycle
The MCP client launches the server, the server uses Python, Python calls photoshop-python-api, and Photoshop responds through COM. If Photoshop is closed, blocked by a dialog, or COM cannot find the right version, the assistant can fail even when MCP configuration looks right.
The first test should stay simple: ask for Photoshop information, create a minimal document, add a text layer, read the layer list, then save into a test folder. Client documents come only after that control.
Available Resources And Tools
Key point: the README exposes three resources and tool families for documents, layers, and session state. The surface is smaller than ComfyUI MCP, but more focused on Photoshop.
Resources
The documented resources are photoshop://info for application information, photoshop://document/info for active document information, and photoshop://document/layers for layers. They are useful before action because they provide starting state.
In a cautious workflow, the assistant reads resources first, confirms the active document, and lists layers. Only then does it create or modify. This sequence reduces wrong-document or wrong-layer mistakes.
Tools
The README groups tools into Document Tools, Layer Tools, and Session Tools. Examples include creating a document, adding a Hello World text layer, opening a PSD and counting layers, or creating a social media post with a gradient background and text.
Those examples are useful but pedagogical. A professional social post still needs typography, brand fit, export settings, dimensions, margins, contrast, and review. MCP automates structure; it does not guarantee taste.
Security, Files, And Limits
Key point: Photoshop MCP can open and save local files, so paths and copies are central. The main risk is altering a working PSD without control.
Files and documents
Creating, opening, and saving documents is powerful. It can overwrite files if the workflow is poorly defined. Use a dedicated output folder, name versions, and work on copies before acting on a client file.
The server reads Photoshop state, but it does not replace visual inspection. Layers can exist and still be misplaced, poorly named, or off brand. Visual control remains mandatory.
Platform limits
The Windows-only limit should not be minimized. COM is Windows-specific. macOS and Linux need another chain. For generative renders or AI workflows, ComfyUI MCP is more portable.
The README says tested Photoshop versions range from CC2017 to 2024, but a real installation may include plugins, IT policies, permissions, or modal dialogs that change behavior.
Comparison With ComfyUI MCP
Key point: Photoshop MCP automates a local Adobe application; ComfyUI MCP orchestrates AI generation graphs.
| Need | Photoshop MCP | ComfyUI MCP |
|---|---|---|
| Open and save PSD | Strong fit | Out of scope |
| Manipulate Photoshop layers | Strong fit | Out of scope |
| Generate images with AI models | Indirect | Strong fit |
| Run on macOS/Linux | No | Yes per README |
| Final creative control | Local Photoshop | ComfyUI graph and outputs |
Combined chain
A coherent workflow can generate assets in ComfyUI and finish them in Photoshop. ComfyUI produces variants, textures, or base images. Photoshop assembles, retouches, adds text, prepares export, and enforces brand.
This separation avoids asking Photoshop MCP for advanced AI generation and ComfyUI MCP for PSD document management.
Educasium Position
Key point: Educasium presents Photoshop MCP as Windows graphics production automation, with mandatory visual validation.
What to teach
A useful exercise starts with a test PSD. The assistant reads resources, lists layers, adds text, creates a background color, saves a copy, and the user verifies in Photoshop.
The skill is state control: which document is open, which layer is active, which file is saved, and which Photoshop version responds. Without that discipline, graphics automation becomes fragile.
The final check includes document size, color profile, margins, text legibility, contrast, layer names, and export format. It also verifies that the saved copy does not overwrite the source PSD and that the destination belongs to the right project. These details look ordinary, but they separate an executed Photoshop action from a usable deliverable. The server accelerates manipulation; validation remains visual and professional. In training, the learner should read state before asking for large changes, because many failures come from acting on the wrong document or an unselected layer.
The creative limit matters too. An assistant can add text, create a simple layer, or save a version, but it does not replace art direction, brand guidelines, image rights, or final review. In a professional workflow, Photoshop MCP should be judged by time saved on repetitive operations, not by a promise of autonomous design. Good practice keeps the source PSD intact, saves dated variants, checks exports, and records which actions were automated.
The exercise should close by comparing the result with the original brief. If the layers exist but the text is unreadable, automation is not finished. If the right file is saved in the wrong folder, it is not finished either. The learner validates result, path, version, and visual quality together.