editmamei/editmamei

**Unlock Photoshop with natural-language photo editing.** AI orchestration, not generation.
> Independent project, not affiliated with or endorsed by Adobe Inc.
Editmamei is a Model Context Protocol (MCP) server that drives desktop Adobe Photoshop. You describe the edit in plain words, your AI assistant plans the steps, and Photoshop carries them out with its standard adjustment layers, masks, selections, and filters. The AI directs and Photoshop edits. No generative model touches your pixels, and no image is uploaded to perform an edit.
It serves photographers and retouchers who want to edit by conversation just as much as developers wiring Photoshop into an AI pipeline. To learn more or get started, visit **editmamei.com**.
**editmamei.com** · Docs · Report a bug
A real run: 73 travel photos, one sentence, graded in Photoshop.
How it works
Editmamei is a pure MCP **stdio** server written in Node. Your AI client launches it as a subprocess and calls its tools over stdio. Each tool builds a small Photoshop script and runs it in your local Photoshop over the platform's native automation channel:
- **Windows** drives Photoshop through COM (the Photoshop.Application object).
- **macOS** drives it through AppleScript / OSA.
- Perception (face, object, and scene detection) runs **on-device** with local ONNX computer vision, so the model gets real document coordinates without your image leaving the machine.
Photoshop does the pixel work with its own engine. Editmamei is the conductor, not a renderer.
Install
Then restart your AI client and ask it: _"Is Photoshop connected?"_
Prefer not to use a terminal? Claude Desktop users can install the one-click .mcpb extension directly (no system Node required, Claude Desktop ships its own runtime). Full walkthrough in the install guide.
**Requirements:** Adobe Photoshop 2026 (v27.x), Windows 10/11 or macOS 13+, Node.js 22+ (for the npm path), and an MCP-compatible AI client (Claude Desktop, Cursor, Claude Code). Editmamei ships no model of its own and runs on whichever client you connect, so every edit spends that client's tokens.
A session, end to end
The discovery chain is ps_ping (liveness), then ps_overview (the working method and capability map), then tools/list (the full schema-typed inventory). After that you just talk to your assistant:
> _"Open my vacation photo."_
> _"Make the sky more dramatic but keep the foreground natural."_
> _"Sharpen the eyes, soften the skin, leave everything else alone."_
> _"Isolate the subject onto its own layer with a clean mask."_
> _"Export a 2048px web JPEG."_
Your assistant composes those into tool calls, checks the result with previews and measurements, and iterates.
Design principles
These are the choices that shape the surface, and the reason an AI assistant can drive it reliably rather than hopefully.
- **Structured results, not prose.** Every tool declares an outputSchema and returns a typed JSON payload. Any tool that changes what is active reports the current document and layer back, so the model keeps an accurate picture of Photoshop's state between calls instead of re-deriving it.
- **Guidance lives in the surface.** ps_overview returns the working method (assess, plan, enact, check, iterate) and ps_list_capabilities returns a live map of what exists. The model orients itself from the server rather than from guesswork or a stale prompt.
- **Measure, don't eyeball.** A dedicated set of verification tools (histogram, region compare, layer-bounds diff, selection preview) lets the model check its own work numerically instead of trusting a thumbnail.
- **On-device perception.** Local computer vision (ONNX) detects faces, objects, and scene regions and returns real document coordinates. The image stays on your machine for this step and is never sent to a cloud vision model.
- **Non-destructive by default.** Adjustments are adjustment layers, and destructive pixel operations run on an auto-created duplicate. An edit is reversible by construction.