OpusGraph

Citation networks treat every paper as an atomic node — you can see which papers cite which, but not what they actually cite at claim level. Two papers citing Zhao 2014 might be citing two different ideas in that paper. Opus Graph uses Claude Opus 4.7's native multimodal PDF reading (no pre-parsing — text, figures, equations, tables fed straight in) to decompose each paper into typed claims (methodological, empirical, theoretical, scoping) and labelled inter-claim relations (supports, depends-on, extends, contradicts). Twenty urban-climate papers become one queryable claim graph: 365 claims, 407 relations, 831 edges across 385 nodes. The headline query walks lineage between papers at claim level. For the anchor pair Sun 2017 → Li 2024, it finds a shared theoretical grounding in Zhao et al. 2014 — not by parsing bibliographies but by traversing the corpus-level claim graph. Three delivery surfaces — a Claude Code skill plugin, an MCP server, and a CLI — all wrap the same Python library. Schema validated against a hand-authored exemplar at 80% lexical recovery; zero out-of-vocabulary labels across the corpus.