Photoshop Mcp

io.github.alisaitteke由 GitHub 账号 alisaitteke 发布v1.7.38更新于 Nov 11, 2026

MCP server for Adobe Photoshop — 128 tools (generative AI + recipes), standalone web UI. Control Ph…

已验证STDIO仅桌面开发者工具媒体与设计

概览

AI 生成的概览

本地 MCP 服务器,让 AI 助手驱动 Adobe Photoshop 编辑图像:去背景、修图、导出、加水印等。

可以做什么
  • 在 Photoshop 中编辑图像:用可编辑蒙版去背景、人像修图、调色、换天空与生成式填充
  • 为网页和社交平台导出:sRGB、锐化、按 Instagram 和 X 尺寸调整
  • 把一张宽幅设计拆成多页轮播图,并批量给整个文件夹加水印
  • 提供 128 个工具(111 个原子工具加 17 个一步式配方),每个配方在 Photoshop 中只算一步撤销
  • 可选择使用内置浏览器聊天窗口,而不必接入 MCP 客户端
适用场景
当你希望用自然语言描述图像修改,并让助手在运行中的 Photoshop 里执行时使用,例如批量修图、生成社交媒体导出变体或拆分轮播图。
运行要求
需要在 Windows 或 macOS 上运行 Photoshop(2012 或更高版本)以及 Node.js 18+,通过 npx 作为本地 stdio 进程启动。可选的独立聊天窗口需要 AI 服务商 API 密钥,或已有的 Claude Code / Gemini CLI 账户;可选的 Jev 路由需要 TYPESAFE_API_KEY。
安装前请注意
该服务器会对 Photoshop 文档执行写入操作,包括批量给文件夹加水印;配方可撤销,但合并和拼合被视为难以撤销。设置 TYPESAFE_API_KEY 后提示词会发送到外部服务,未设置则不发送。这是非官方社区项目,与 Adobe 无关,生成式填充需要 Adobe 账户。

安装

在 SourceWeft 中

  1. 打开 控制台中的 Photoshop Mcp,将其添加到工作区。
  2. 为需要使用其工具的对话启用该服务。

Desktop only,通过 STDIO。 STDIO 服务会启动本地进程,因此需要 SourceWeft 桌面宿主。

其他 MCP 客户端

参照 仓库 中的启动说明。

README

Photoshop MCP

[Photoshop MCP — tell Photoshop what you want, AI does the clicking]

Languages: English · 简体中文 · Español · Deutsch · 日本語 · Türkçe · Website

[npm version] [GitHub release] [Action Plan] [Jev routing] [License: MIT] [TypeScript] [Platform] [MCP Registry] [MCP Toplist] [Website]

[Photoshop MCP Server MCP server – quality and maintenance score on Glama]

Chat with Photoshop like a colleague. Describe what you want in plain words — "remove this background", "resize these for Instagram" — and your AI assistant does the clicking for you. Works with Cursor, Claude, or the built-in chat window. No code, no scripts, no IDE required.

Note: This is an unofficial, community-maintained project and is not affiliated with or endorsed by Adobe Inc.

New: instant edits with Jev

[Jev routes each prompt before any LLM runs: one safe command goes straight to Photoshop with no LLM call; plans, visual checks and vague requests go to the LLM]

Not every prompt needs a language model. The built-in chat window can now ask Jev, TypeSafe AI's System One model, where each message should go before any LLM runs. When Jev is confident you asked for one of 26 safe commands and recipes (undo, opacity, blend mode, remove background, color grade, split carousel, …), or a short chain of them like "black and white, then opacity 50", it runs in Photoshop straight away: no LLM call, no LLM tokens. Anything else, and hard-to-undo steps like merge or flatten, goes to the Action Plan.

bash
TYPESAFE_API_KEY=... npx -p @alisaitteke/photoshop-mcp ui

Experimental and opt-in. Without the key the UI works exactly as before and nothing is sent to TypeSafe; with it, prompts also go to api.typesafe.ai. Routes, thresholds and the off switch: docs/standalone-ui.md.

What can it do?

  • ✂️ Remove backgrounds — subject isolated with a clean, editable mask
  • 👤 Retouch portraits — skin smoothing, tone fixes, dodge & burn setup
  • 🌐 Export for web & social — sRGB, sharpened, correctly sized for Instagram, X, and more
  • 🎞️ Make carousels — split one wide design into seamless, numbered slides
  • 💧 Watermark in bulk — a whole folder of photos in one go, originals untouched
  • 🎨 Color grade & more — film looks, sky replacement, generative fill (Adobe account required)
  • ⏪ Stay safe — every multi-step "recipe" is a single undo step in Photoshop

Under the hood: 128 tools (111 atomic + 17 one-step recipes) — full list in docs/available-tools.md.

Try saying

Remove the background from this portrait — keep it editable with a mask.
[Remove background: subject isolated on transparency]
Enhance this portrait — smooth the skin and fix the tones, medium intensity.
[Enhance portrait: skin smoothing + auto tone in one undoable step]
Prepare this design for web, then export Instagram and X post variants.
[Prepare for web: sRGB, downscale, sharpen, optimized JPEG]
Split this wide banner into a 5-slide seamless Instagram carousel.
[Split one wide document into numbered carousel slides]

More recipes (batch watermark, passport photos, CSV-driven cards, mockups, …) and pre-engineered prompt templates: docs/prompt-layer.md.

Get started

You need Photoshop running (Windows or macOS, any version 2012+) and Node.js 18+.

Option 1 — Easiest: the built-in chat window

bash
npx -p @alisaitteke/photoshop-mcp ui

A chat window opens in your browser. Sign in with an AI provider API key — or reuse your existing Claude Code / Gemini CLI account, no key needed.

[Standalone UI Screenshot]

Details, providers, Action Plan (API key or CLI account), and security notes: docs/standalone-ui.md.

Option 2 — Inside your AI app (Cursor, Claude, VS Code)

Cursor (shows the Photoshop logo in the MCP list): install the plugin from Customize → Plugins, or search the Cursor Marketplace for photoshop-mcp once it is listed. That package still launches npx -y @alisaitteke/photoshop-mcp; the plugin manifest is what supplies the icon.

Until the Marketplace listing is live, copy .cursor-plugin/, mcp.json, and assets/ into ~/.cursor/plugins/local/photoshop-mcp and reload the window — details in CONTRIBUTING.md. The Install-in-Cursor button and a raw mcp.json entry still work, but they show a generic icon.

[Install in Cursor] [Install in VS Code]

Claude Code:

bash
claude mcp add photoshop -- npx -y @alisaitteke/photoshop-mcp

Or add this to your MCP client's config (Cursor, Claude Desktop, …):

json
{  "mcpServers": {    "photoshop": {      "command": "npx",      "args": ["-y", "@alisaitteke/photoshop-mcp"]    }  }}

How it works

  1. You type what you want in plain language.
  2. The AI plans the steps, checking the document state first.
  3. Photoshop executes — each recipe lands as one undoable step.

Something went wrong? The AI reads the structured error and knows what to try next. Common fixes: docs/troubleshooting.md.

Documentation

Contributing

Contributions are welcome! Please read CONTRIBUTING.md before opening a PR.

❤️ Contributors & Credits

Thanks to all our contributors!

[Contributors]

Maintainer

Built by Ali Sait Teke — GitHub · LinkedIn.

License

MIT

Anonymous, aggregated usage analytics are collected by default and can be disabled anytime — details in docs/anonymous-usage-analytics.md.

来源:README.md,提交 d25cba0

工具

0
SourceWeft 尚未在沙箱中启动过该软件包,因此其工具未知。

版本历史

17
  1. v1.7.38最新Nov 11, 2026
  2. v1.7.37Oct 15, 2026
  3. v1.7.34Oct 8, 2026
  4. v1.7.33Oct 5, 2026
  5. v1.7.32Oct 1, 2026
  6. v1.7.30Oct 1, 2026
  7. v1.7.28Oct 1, 2026
  8. v1.7.27Sep 30, 2026
  9. v1.7.25Sep 30, 2026
  10. v1.7.24Sep 28, 2026
  11. v1.7.22Sep 27, 2026
  12. v1.7.21Sep 25, 2026
  13. v1.7.20Sep 22, 2026
  14. v1.7.19Sep 22, 2026
  15. v1.7.18Sep 21, 2026
  16. v1.7.15Sep 16, 2026
  17. v1.7.14Sep 16, 2026

评分

已满足适用检查项的 50%

该评分衡量此收录条目的完整程度,依据注册中心条目、端点与公开来源计算得出。它不是对服务器本身的评价,也不表示服务器安全或当前可用:可连通性是带日期的检查结果,评分也不包含安装量或用户评价。

  • 近 30 天内可连通25 分

    请保持端点能够响应 MCP 握手;SourceWeft 每 30 天探测一次。

  • 已声明认证方式10 分

    请在注册中心条目中声明凭据请求头,或保持端点能够响应探测,以便判定其认证方式。

  • 已列出工具15 分

    请让服务器在无凭据时响应 tools/list,或发布可在沙箱中无凭据启动的软件包。

  • 描述必需15 分
  • 图标10 分
  • 文档15 分
  • 许可证10 分