Labsco
MCP SERVER

paperbanana

by llmsresearch

Turn a methodology write-up or a CSV of results into a publication-quality figure, then let a critic agent look at the image and revise it.

Image Generation
Summary
A figure that gets looked at and revised before you ever see it.

The loop that matters is visualizer-to-critic: the generated image is read back against the source text and the caption, and the description is rewritten from what is actually wrong with the picture.

What it is

An agentic pipeline for academic illustration. A planner drafts the diagram from your text, a stylist applies venue guidelines, a visualizer renders it, and a critic reads the rendered image back and rewrites the description — repeating for a set number of rounds or until the critic is satisfied.

What you get
  • Eleven MCP tools: `generate_diagram`, `generate_plot`, `continue_run`, `continue_diagram`, `continue_plot`, `evaluate_diagram`, `evaluate_plot`, `orchestrate_figures`, `batch_diagrams`, `batch_plots` and `download_references`
  • Methodology diagrams from a text file or a PDF with per-page selection, optionally guided by a hand-drawn sketch or a prior figure
  • Statistical plots rendered as VLM-generated matplotlib code — no image-generation provider or credentials needed for that path
  • Whole-paper figure packages via `orchestrate_figures`: the paper is parsed, figures are planned from its section structure, and the output folder carries `figure_package.json`, `figures/`, `figures.tex` and `captions.md`
  • Batch runs from a YAML or JSON manifest, with an optional `composite` section that stitches the panels into one labelled figure
  • Comparative evaluation against a human reference on four dimensions — faithfulness and readability as primary, conciseness and aesthetics as secondary
  • Venue style packs for `neurips`, `icml`, `acl` and `ieee`, with room for your own under `~/.config/paperbanana/venues/`
Requirements

Python 3.10+ and an API key: an OpenAI key, an Azure OpenAI / Foundry endpoint via `OPENAI_BASE_URL`, or a Google Gemini key. `pip install paperbanana`, or run the MCP server without a clone using `uvx --from paperbanana[mcp] paperbanana-mcp` with `GOOGLE_API_KEY` in the env. The local web UI needs `pip install 'paperbanana[studio]'` and PDF input needs `paperbanana[pdf]`. `paperbanana polish` requires an image provider that supports guided edits. Note this is a community implementation of a published paper, not affiliated with the original authors, so it may differ from the system described there.