What it does Feed Paper2Agent a GitHub URL and it clones the repo, hunts down tutorials, executes them, extracts working tools, and wraps everything in an MCP server for Claude Code or Gemini CLI. The output is a domain-specific agent—say, an AlphaGenome genomics assistant or a Scanpy single-cell preprocessor—that answers natural-language questions using the paper’s actual code.

The interesting bit The pipeline doesn’t just dump code into a prompt. It runs the notebooks, generates tests, checks coverage with pytest, and lints the extracted tools before minting the MCP server. That five-step verification is the guardrail; without it you’re just hoping a LLM hallucinated the right API calls.

Key highlights

• One-shot creation: bash Paper2Agent.sh --github_url <url> and wait 30 min to 3+ hours
• Auto-discovers and filters tutorials by title or URL; supports private repos with --api
• Emits isolated Python environments, coverage reports, and pylint scores—not just vibes
• Ships with three live remote MCP servers (AlphaGenome, Scanpy, TISSUE) hosted on Hugging Face
• Optional --benchmark flag extracts benchmark questions and scores the final agent

Caveats

• Requires Claude Code (or Gemini CLI) and an Anthropic API budget; the README quotes ~$15 for a complex repo like AlphaGenome using Claude Sonnet 4
• Processing time is measured in hours, not seconds, and the “own base agent” is still listed as “coming soon”
• Jupyter Notebook repo with 2,231 stars but no tagged topics, so discoverability is currently word-of-mouth

Verdict Worth a spin if you’re a bioinformatician or ML researcher who lives in Claude Code and wants to stop copy-pasting from Methods sections. Skip it if you need instant tooling or lack the API budget for a multi-hour autonomous pipeline.