// the find
Open-Source-Legal/OpenContracts
The open document intelligence platform for builders and hackers - DMS for the agentic world
OpenContracts is a self-hosted document management system built around a citation graph — documents are nodes, citations are edges, and both humans and AI agents contribute to building those edges over time. It ships a GraphQL/REST API, a React annotation UI, an MCP server, Celery-backed async processing, and structured extraction all in one stack. The target is teams that need to reason across a repository of documents rather than just store them.
The citation graph model is the real idea here — tracking cross-references as first-class edges rather than buried text means agents can actually traverse related documents instead of hallucinating connections. The MCP server is a practical addition: expose a corpus to Claude or Cursor with zero glue code, and the tools cover search, annotation listing, and relationship traversal. The pipeline abstraction is clean — parsers, embedders, and thumbnailers are swappable components with defined interfaces, so plugging in a custom parser doesn't require forking the annotation layer. PDF coordinate mapping via PAWLS is genuinely hard to get right, and having layout-faithful annotations that survive round-trips to the database is worth more than it sounds.
The operational footprint is heavy: the local stack is Django, Celery beat, Celery worker, Flower, Postgres, Redis, and a separate Docling microservice for PDF parsing — that's a lot to keep running for what might start as a small team's document library. The project is still in v3.0.0.b3 and the 100+ changelog fragments from the past weeks suggest active churn in core areas (permissions, enrichment, SSRF hardening, bulk delete) that will bite early adopters. Format support is thin — PDF, DOCX, and plain text only; no Markdown, HTML, or ePub, which limits the usecase outside legal and research documents. The 'shared graphs compound' community angle is pitched as OpenStreetMap for citations, but without an actual network of public corpuses, you're self-hosting an expensive stack to get a benefit that only materializes if others join.