v0.1.0 and v0.1.1: PyPI, MCP server, Claude Code skills, and MCP Registry

[dippatel1994]

PaperBanana is now installable from PyPI and registered on the Official MCP Registry. Here's what shipped and what's next.

What's available

Three install paths, depending on how you want to use it:

pip install paperbanana gives you the CLI and Python API. Generate diagrams with paperbanana generate --input method.txt --caption "...", generate plots from CSV/JSON, and run VLM-as-a-Judge evaluation against human references.

pip install paperbanana[mcp] adds the MCP server. If you're using Cursor, Claude Code, VS Code with Copilot, or any other MCP-compatible client, you can call PaperBanana's tools directly from your editor.

uvx --from "paperbanana[mcp]" paperbanana-mcp runs the MCP server without installing anything locally. One JSON block in your editor config and you're set. See the [MCP Server Setup](https://github.com/llmsresearch/paperbanana/wiki/MCP-Server-Setup) wiki page for copy-paste configs for each client.

Claude Code skills ship with the repo. Clone the repo, open Claude Code in that directory, and you get three slash commands: /generate-diagram, /generate-plot, /evaluate-diagram. These are project-scoped and call the MCP tools under the hood.

MCP Registry listing. PaperBanana is published on the [Official MCP Registry](https://registry.modelcontextprotocol.io) and submitted to [mcp.so](https://mcp.so). This means MCP clients that pull from the registry can discover it directly.

What changed between v0.1.0 and v0.1.1

v0.1.0 was the initial PyPI release with everything above. v0.1.1 added server.json metadata for the Official MCP Registry and the registry submission itself. No functional changes between the two.

Current state of the reference dataset

This is the most important thing to know about PaperBanana right now. The paper by Zhu et al. uses 292 curated reference examples. We have 13. The Planner agent uses retrieved references as few-shot demonstrations, so the quality and diversity of the reference set directly bounds output quality.

We're actively looking for contributions here. If you've come across a paper with a clear methodology diagram, the easiest way to help is to open a [Reference Example issue](https://github.com/llmsresearch/paperbanana/issues/new?template=reference_example.yml) with the arXiv link, figure number, and category. We handle extraction and curation. If you want to submit a complete parsed tuple via PR, the [Adding Reference Examples](https://github.com/llmsresearch/paperbanana/wiki/Adding-Reference-Examples) wiki page has the full guide.

Categories most needed: Science & Applications and Vision & Perception.

What's next

Expanding the reference dataset is the priority. Beyond that, the next priority is to improve agentic architecture to support more datasets we will support, so at inference time we retrieve closes examples to the asked paper to improve performance. We can also tap into evaluating the solution and share our findings. If you're interested in contributing to either of these, check the [Contributing](https://github.com/llmsresearch/paperbanana/wiki/Contributing) guide or open an issue to coordinate.

Full changelog in the [wiki](https://github.com/llmsresearch/paperbanana/wiki/Changelog). Release assets on the [v0.1.0](https://github.com/llmsresearch/paperbanana/releases/tag/v0.1.0) and [v0.1.1](https://github.com/llmsresearch/paperbanana/releases/tag/v0.1.1) pages.