Paper2Agent: research papers as interactive AI agents

What it adds over the summary

The summary establishes what Paper2Agent does (multi-agent conversion of papers into tested MCP servers, shipped as a skill, three demos). The page adds the concrete install/invocation mechanics, credential handling, hosted endpoints, and publication venue — the operational detail an adopter needs.

How to use it

Install the skill by cloning and copying the full skill folder into your host's personal skill location:

git clone https://github.com/jmiao24/Paper2Agent.git
cd Paper2Agent

Claude Code:

mkdir -p "$HOME/.claude/skills/paper2agent"
cp -R skills/paper2agent/. "$HOME/.claude/skills/paper2agent/"

Codex:

mkdir -p "$HOME/.agents/skills/paper2agent"
cp -R skills/paper2agent/. "$HOME/.agents/skills/paper2agent/"

Restart the coding agent if the skill does not appear. Conversion is invoked via /paper2agent in Claude Code or $paper2agent in Codex, e.g.:

/paper2agent Convert https://github.com/scverse/scanpy into tested MCP tools in Scanpy_Agent.
Focus on the "Preprocessing and clustering" tutorial.

You can scope conversion to specific tasks, a tutorial title, or a tutorial notebook URL. Output is a delivered ZIP whose USAGE.md records the tested interpreter, server entry point, required env vars, and supported platforms.

Credentials: API keys are read from a named environment variable (e.g. ALPHAGENOME_API_KEY) that you pass to the agent; the page states credentials stay outside generated code, notebooks, reports, and the ZIP.

Connecting a generated/hosted server to Claude Code over HTTP:

claude mcp add --transport http <MCP_NAME> <MCP_ENDPOINT_URL>

Verify with claude mcp list or /mcp. Three hosted servers exist:

• AlphaGenome: https://Paper2Agent-alphagenome-mcp.hf.space
• Scanpy: https://Paper2Agent-scanpy-mcp.hf.space
• TISSUE: https://Paper2Agent-tissue-mcp.hf.space

Requirements

Host must support skills, shell access, and parallel subagent spawning (the coordinator launches specialist and fresh verifier agents through the host). Runtime needs Python and Git, plus any R, native CLI, data, API, or GPU dependencies of the target repo; the skill builds isolated project environments and pins tested versions.

Worth noting

The work is published in Nature (Miao et al., 2026, doi:10.1038/s41586-026-11044-y) — a signal of vetting beyond a typical GitHub release. The page is documentation, not evaluation: it gives no numbers on conversion success rate, tool correctness, or how often verification passes, so an adopter should treat the three curated demos as existence proofs rather than a benchmark.