Imgx

What imgx does

The imgx MCP server connects an AI coding environment to image generation and editing workflows. Agents can create an image from a text prompt, modify an existing file with text instructions, or continue editing the most recent result without supplying its path again.

Each generation starts an editing session. Later changes remain attached to that session, allowing an agent to build an edit chain rather than treating every request as an unrelated operation. The server also exposes project history, session switching, history cleanup, and output-directory management.

The repository describes an accompanying image-generation skill for Claude Code and Claude Desktop. That skill is separate from the MCP server: it provides guidance for constructing prompts, choosing model-related options, selecting platform-aware dimensions, and applying supported editing approaches. The MCP tools themselves can be used by any MCP-compatible client.

How it works

The imgx MCP server communicates over the standard MCP configuration used by supported clients. A typical configuration starts the package with npx and passes a provider API key through the environment. On Windows, the README uses cmd with /c instead of invoking npx directly.

Generated files and history are organized around the detected project root. When a project root is available, .imgx/ contains the output and history data, including output-history.json. Without a detected root, images are placed under ~/Pictures/imgx/, while global history is stored under ~/.config/imgx/output-history.json. Clients using the same project root share its history.

edit_last follows the current position in a session. Undoing an edit moves the session backward; redoing restores a previously undone step. Creating a new edit after undo branches from the current position, and abandoned entries and their files are removed. File paths are returned in tool responses, which also include inline image previews as base64 data.

Setup and configuration

Add an imgx entry to the MCP configuration used by the client:

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

Configure at least one image provider before making requests. The documented quick-start example uses GEMINI_API_KEY; the repository directs users to Google AI Studio to obtain that key. The set_output_dir tool can change where image files are written, and an option can move existing files. History remains in the project or global history location rather than moving with the image output directory.

Claude Code users can optionally copy the package’s image-generation skill into .claude/skills/. Claude Desktop users can upload the provided skill ZIP through its Skills settings. The skill is not required to expose or call the MCP tools.

Tools and capabilities

The server exposes these tools:

• generate_image: Create an image from a text prompt.
• edit_image: Modify an existing image using text instructions.
• edit_last: Modify the latest image in the active session without an input path.
• undo_edit and redo_edit: Move backward or forward through the active edit chain.
• edit_history: Display sessions and edit metadata.
• switch_session: Select another session for continued work.
• clear_history: Remove project history, optionally deleting image files.
• set_output_dir: Change the default image output location, optionally moving existing files.
• list_providers: Show available providers and their capabilities.

Limitations and notes

The imgx MCP server depends on a configured image provider and its credentials. The README’s documented setup shows Gemini, while provider discovery is available through list_providers. Each generation creates a new session, and subsequent edit_last operations use that session’s output directory. Skill-based prompt guidance is documented for Claude Code and Claude Desktop; other clients can still use the MCP tools directly but do not receive that skill automatically.