Claude Paper

Claude Paper is a DeepSeek Harness plugin in the broad Agent Skills sense rather than a Cordis Bundle. Its pinned README describes one research-paper workflow shared by Claude Code, Codex, OpenCode, and DeepSeek Harness. A user can ask an agent to summarize or study a PDF, a direct PDF URL, or an arXiv abstract URL. The project then keeps the source PDF, complete extracted text, metadata, generated learning material, figures, and optional code work together in a local paper library. That makes it a candidate for people who want research work to leave inspectable files rather than only a transient chat answer.

The quick route is intentionally smaller than a full study: the README says it produces a 300–500 word overview for screening a paper before committing more context and compute. The deeper route parses the document, assesses complexity and type, prepares a README, summary, insights, questions, and sometimes methodology or reflection material. It can also look for code links and generate a clean reference implementation or notebook. These are project claims recorded from the source, not a claim that every PDF will parse cleanly or that generated explanations are academically correct.

Its local viewer is a separate practical reason to use it. The pinned documentation says the web interface opens on localhost:5815 and can browse a paper collection, search it, render mathematics, and expose generated notebooks and demonstrations. The storage layout is explicit: paper folders live under ~/claude-papers/papers and the global index is ~/claude-papers/index.json. This transparency is useful when a study needs later review, but it also means PDFs and their extracted content are retained on the workstation until the user manages them.

Claude Paper should not be mistaken for a hosted literature database, an official DeepSeek product, or a guarantee of citation quality. It is a community-maintained cross-agent toolkit. The package manifest pins its own release at 1.2.1 and requires Node 20.19 or 22.12+, while the project documentation additionally calls out poppler-utils for figure extraction. No DSH-specific version range appears in the reviewed package. Evaluate it first with non-sensitive papers and verify both the host Skill discovery and the generated local files before relying on it for research deliverables.