Replication Radar MCP Server

Replication Radar

🔗 Live demo → https://openaire-hackathon.netlify.app

▶ Watch the 3-min demo → https://youtu.be/hVyLafY3Y3E · reproduce it step by step → https://openaire-hackathon.netlify.app/demo.html

A tool that makes the OpenAIRE Graph more useful for replication. Search a research field and it answers the question the Graph structurally cannot: what high-impact work is worth replicating, has it already been independently checked — with what verdict — and is the software reusable?

Ships as a live web app (the link above — pure static, queries OpenAIRE + the nanopub network + GitHub/Software Heritage from the browser) and an MCP server (this package) that exposes the same engine to any agent. Built for the OpenAIRE AI Hackathon (Theme B), CC-BY.

OpenAIRE's only value signal is citation-popularity (BIP! influence / popularity / impulse, classes C1–C5) — paper-bound, and orthogonal to whether a claim is true. The Radar joins three sources to add a replication layer on top:

• OpenAIRE Graph — impact-ranks candidate papers (api.openaire.eu/graph/v1).
• Software Heritage + repo signals — surfaces reusable method software.
• Science Live nanopub verdicts — the "already checked → did it hold" overlay.

OpenAIRE AI Hackathon · Theme B (Build) · CC-BY. Built to be reused through the forrt-replication-template: discovery at the start of a replication, where the template's existing skills handle the nanopub chain at the end.

Tools

The verdict tools pull live from the nanopub network (the FORRT Outcome/CiTO templates on query.knowledgepixels.com); the bundled verdicts.json is an offline fallback. So the MCP is the verified-knowledge layer — pair it with the OpenAIRE MCP and an agent has both the structural Graph and "has this been checked, and did it hold".

The reproduction-vs-replication distinction, made computable

A reproduction re-runs the original code; a replication tests the same claim by a different route. So the Radar filters tooling by author-disjointness from the original paper — e.g. for Phillips et al. 2009, the dismo package (co-authored by Phillips & Elith) is flagged rooted / non-independent, while biomod2 and jSDM are independent. That filter is the difference between the two, and it's the thing that makes this replication-aware rather than just "find the code".

Run

pip install -e .                       # installs the `mcp` runtime
python -m replication_radar.server     # stdio MCP server

Add to an MCP client (.mcp.json):

{ "mcpServers": {
  "replication-radar": { "command": "python", "args": ["-m", "replication_radar.server"] }
} }

The core (OpenAIRE client + radar logic) is stdlib-only — try it without the MCP runtime:

PYTHONPATH=src python3 demo_sdm.py     # live vertical-slice demo on SDM

Configuration

Known limits (v1, honest)

• Keyword-bound discovery. OpenAIRE free-text terms are AND-ed; long queries return nothing. Use short topics. The VERIFIED overlay is guaranteed (resolved from the verdict index directly), but OPEN-target recall depends on the query.
• No graph-relation traversal on the public API (paper→its software/data/grant edges aren't exposed): tooling/data are matched heuristically by topic + author independence, not by a hard relation. Upgrades cleanly if a gateway exposes relations.
• Funder context is field-level, not per-paper (per-paper funder attribution is not reachable); budgets are frequently reported as 0 in records.
• The verdict index ships 6 source works / 12 chains (Science Live). Extend data/verdicts.json to grow coverage.