imgx-mcp

AI image generation and editing MCP server. Works with Claude Code, Gemini CLI, Cursor, Windsurf, and any MCP-compatible tool.

Generate images from text, edit existing images with text instructions, iterate on results — all from your AI coding environment.

What sets imgx-mcp apart

• No prompt engineering — Your AI agent keeps conversation context and auto-constructs optimized prompts. Say what you need; the agent handles prompt structure, model selection, and platform-specific sizing
• 24 editing techniques built in — Atmosphere, composition, style transfer, element manipulation, and trending styles — bundled as a Skill your agent applies on demand
• Session management with undo/redo — Edit iteratively, step back to any point, branch off, or switch between parallel sessions — version control for images

Quick start

Add to your tool's MCP config (.mcp.json, settings.json, etc.):

{
  "mcpServers": {
    "imgx": {
      "command": "npx",
      "args": ["--package=imgx-mcp", "-y", "imgx-mcp"],
      "env": { "GEMINI_API_KEY": "your-key" }
    }
  }
}

That's it. Your AI agent can now generate and edit images.

Windows: Replace "command": "npx" with "command": "cmd" and prepend "/c" to the args array.

Skill (Claude Code)

For Claude Code users, imgx-mcp includes an image-generation skill — a guided prompt that teaches Claude how to use the MCP tools effectively. With the skill installed, type /image-generation to start a guided workflow.

Install the skill

Copy the skill directory from the npm package or GitHub repository to your project:

# From npm (after npx has cached the package)
cp -r $(npm root -g)/imgx-mcp/skills .claude/skills

# Or from the GitHub repository
curl -sL https://raw.githubusercontent.com/somacoffeekyoto/imgx-mcp/main/skills/image-generation/SKILL.md \
  -o .claude/skills/image-generation/SKILL.md --create-dirs
curl -sL https://raw.githubusercontent.com/somacoffeekyoto/imgx-mcp/main/skills/image-generation/references/providers.md \
  -o .claude/skills/image-generation/references/providers.md --create-dirs

Or place skill files manually:

your-project/
  .mcp.json                              ← MCP server config (Quick start above)
  .claude/
    skills/
      image-generation/
        SKILL.md                         ← skill prompt
        references/
          providers.md                   ← provider reference

The skill files are included in the npm package under skills/ and in the GitHub repository.

Personal skill (all projects): Place in ~/.claude/skills/image-generation/ instead of .claude/skills/.

Claude Desktop

Claude Desktop supports skills via ZIP upload:

1. Download image-generation-skill.zip from the repository (or find it in the npm package under dist/)
2. In Claude Desktop: Settings > Profile > Customize > Skills > Add Skill
3. Upload the ZIP

Update the skill by re-downloading and re-uploading the ZIP after new releases.

What the Skill brings

The MCP server gives the AI the ability to generate and edit images. The Skill adds the knowledge of how to use those tools well — so you don't need to learn prompt syntax, model specifications, or service-specific parameters.

• Automatic prompt construction — Say "I need a cover image." The AI builds a structured prompt using the Subject-Context-Style framework: what to show, where to place it, how it should look
• 24 editing techniques — Atmosphere adjustment, composition changes, element manipulation, style transfer. "Make it warmer" or "add depth of field" — the AI selects the right instruction for the model
• Intelligent model selection — Starts with the free model. Suggests paid upgrades only when your needs exceed free tier capabilities, and explains what changes
• Platform-aware sizing — "Twitter OGP" or "App Store screenshot" — the AI picks the correct aspect ratio and resolution. Covers social media, OGP, app stores, print, and blog platforms
• Trending style templates — Ghibli, action figure in box, 3D clay, pixel art, chibi, and more. Name the style and the AI applies the right prompt structure
• Multi-image consistency — Design tokens and character DNA templates maintain visual coherence across slide decks, social media series, and brand assets

The image generation models already have these capabilities. The Skill is what makes them accessible without specialized knowledge.

MCP server vs Skill

Recommended: Set up the MCP server (Quick start) + install the skill if you use Claude Code.

MCP tools

The .imgx/ directory holds both edit history and default image output. Its location depends on project root detection:

All clients that resolve to the same project root share the same history. Each session gets its own subdirectory. File paths are returned in the response. Inline image preview is included in MCP responses (base64).

Iterative editing

The edit_last tool uses the output of the previous generate_image or edit_image call as input. This enables a conversational workflow:

"Generate a coffee shop interior" → generate_image
"Make the lighting warmer"        → edit_last
"Add a person reading a book"     → edit_last

No need to specify file paths between steps.

Session management

Each generate_image call starts a new session. Subsequent edit_last calls are added to the same session, forming an edit chain. Each session has its own output directory.

Undo / Redo — Step backward and forward through the edit chain:

generate → edit_last → edit_last → edit_last
                                    ↑ current
                       ← undo_edit
                       ↑ current
                            redo_edit →
                                    ↑ current

After undo, calling edit_last branches from the current position (abandoned entries and their files are deleted from disk).

File naming — edit_last generates sequential filenames based on the origin file:

generate_image             → cover.png
edit_last                  → cover-1.png
edit_last                  → cover-2.png

generate_image (no output) → imgx-a1b2c3d4.png
edit_last                  → imgx-a1b2c3d4-1.png

Session switching — Use edit_history to see all sessions, then switch_session to resume a previous session. The edit_last tool will use the current position in the switched session.

Output directory — edit_last inherits the output directory from the session. If generate_image was called with output_dir, all subsequent edit_last calls in that session output to the same directory. The output_dir path is recorded as session metadata in output-history.json. This only affects where image files are saved — history always stays in .imgx/ (or the global config directory).

API key setup

Set up at least one provider:

Gemini — get a key from Google AI Studio (free tier available for gemini-2.5-flash-image):

imgx config set api-key YOUR_GEMINI_API_KEY --provider gemini

OpenAI — get a key from OpenAI Platform:

imgx config set api-key YOUR_OPENAI_API_KEY --provider openai

Keys are stored in ~/.config/imgx/config.json (Linux/macOS) or %APPDATA%\imgx\config.json (Windows). Alternatively, pass keys via the env section in your MCP config, or set environment variables:

export GEMINI_API_KEY="your-api-key"
export OPENAI_API_KEY="your-api-key"

Only include the API keys for providers you want to use. At least one is required.

MCP configuration by tool

Claude Code

.mcp.json in your project root:

{
  "mcpServers": {
    "imgx": {
      "command": "npx",
      "args": ["--package=imgx-mcp", "-y", "imgx-mcp"],
      "env": { "GEMINI_API_KEY": "your-key", "OPENAI_API_KEY": "your-key" }
    }
  }
}

Gemini CLI