alisaitteke/photoshop-mcp

What it solves

Photoshop MCP solves the difficulty of controlling Adobe Photoshop via AI assistants. Standard automation scripts are often brittle and lack context, causing AI agents to waste tokens on trial-and-error or leave documents in inconsistent states. This project provides a robust bridge that gives AI assistants state awareness and a structured way to execute complex design workflows using natural language.

How it works

It implements a Model Context Protocol (MCP) server that allows LLMs (like Claude or Cursor) to programmatically control Photoshop. The system features:

• State Awareness: Tools like get_state and get_preview allow the AI to see the current document and verify results via base64 JPEGs.
• Recipe Tools: 16 high-level "recipes" (e.g., background removal, color grading) that wrap multi-step processes into a single, undoable action.
• Prompt Layer: 23 pre-engineered templates that help LLMs translate vague user requests into precise tool calls.
• Standalone UI: A local web interface that supports API keys or existing CLI subscriptions (Claude Code/Gemini CLI) to drive Photoshop without needing an IDE.
• Action Plan (beta): A mode that plans an entire sequence of steps in one LLM call and executes them in a single pass to reduce round-trips.

Who it’s for

Designers and developers who want to automate repetitive Photoshop tasks, create data-driven graphics, or perform complex image manipulations using natural language commands.

Highlights

• 118 Total Tools: Includes 102 atomic tools and 16 outcome-oriented recipes.
• Generative AI Integration: Supports Adobe Firefly features like generative fill, expand, and upscale.
• Neural Filter Support: Access to skin smoothing, depth blur, and colorization via an optional UXP bridge plugin.
• Non-Destructive Workflows: Recipes are designed to be single undo steps to preserve original pixels.
• Flexible Auth: Supports multiple providers (Anthropic, OpenAI, Google, OpenRouter) via API keys or CLI OAuth sessions.