mcp-gpt-image

# 🎨 mcp-gpt-image **An image MCP server built for the second draft, not the first.** Getting one picture out of a model is the easy part. The afternoon goes on the twelve that come after it: the sky is too bright, the logo drifted, the text on the sign came out wrong. Most image connectors stop at the first picture and leave you to redo the loop by hand. Three tools, OpenAI's [gpt-image-2.5](https://developers.openai.com/api/docs/guides/image-generation) in both of its variants, and file paths instead of base64. ## 🧱 The problem You write a prompt, you get a picture, you want the sky darker. So you write the whole prompt again with "darker sky" appended, because the model kept nothing from the first call. Every correction is a full rewrite, and the model has no idea what it just drew. When the connector hands images back as base64 it gets worse. The picture is now sitting in the conversation: it cost tokens to arrive, it costs tokens again on every turn it stays there, and feeding it into the next step means sending those same bytes back up. Two rounds of that on a 1536x1024 image and the context window is mostly pixels. This server takes the other route on both counts. Images travel as **file paths**: `generate_image` writes a file and returns where it is, `edit_image` and `refine_image` take paths as input, so a picture moves from one step to the next without ever entering the conversation. And `refine_image` **holds the thread** with the model, so turn one is a full description and turn two is "make the sky darker". Nothing is re-uploaded, nothing is re-described. ## 📦 Install ```bash claude mcp add gpt-image -e OPENAI_API_KEY=sk-... -- uvx mcp-gpt-image ``` Or in any MCP client's config: ```json { "mcpServers": { "gpt-image": { "command": "uvx", "args": ["mcp-gpt-image"], "env": { "OPENAI_API_KEY": "sk-...", "GPT_IMAGE_OUTPUT_DIR": "/home/you/images" } } } } ``` Needs Python 3.12 or newer, [uv](…