Paperbanana MCP Server

Research figure workflows in PaperBanana start with methodology text or structured plotting data. The generate_diagram tool pairs a paper excerpt or methodology description with a caption describing the intended message, then returns a PNG diagram. For statistical graphics, generate_plot takes data as a JSON string and a description of the desired plot, also returning PNG output. Both generation tools accept a refinement iteration setting. Diagram generation combines reference retrieval, planning, styling, rendering, and critique; plots are rendered through generated matplotlib code.

To assess a diagram, evaluate_diagram uses paths to the generated image and a human-drawn reference, together with the original methodology context and caption. It reports scores for faithfulness, conciseness, readability, and aesthetics, plus an overall winner. The Python MCP interface uses stdio for IDE integration; the project also offers a CLI, Python API, and local Gradio interface. PaperBanana is an unofficial community implementation based on the published research, without affiliation with or endorsement from Google Research or the original authors.