The curated DeepSeek Harness
research plugins

Search: one cross-database query with de-duplication, or a single source — Semantic Scholar, PubMed, Google Scholar, arXiv, bioRxiv/medRxiv. Follow a citation graph, look up authors, ask for recommendations. Read: resolve a DOI to its PDF and extract the full text in slices, so the agent works from the paper rather than the abstract; open-access copies work anywhere, paywalled publishers work on a network with institutional access. Cite: hand it a draft with numbered claims and get real references back. Draw: generate and edit scientific figures. An API-key card lives in Settings, and every billed answer ends with what it spent and what is left; arXiv, bioRxiv and DOI full text cost nothing.

The conversion runs on your host through poppler: pdftotext -bbox gives word geometry, from which it rebuilds columns by detecting the gutter, reassembles full-width lines, normalises headings, recovers tables from column geometry, and keeps sub/superscripts and dot-leader tables of contents. Figures and display equations are cropped with pdftoppm and pulled with pdfimages, and the equation text is kept as a hidden span so the page stays searchable. The reader opens as a real third column when the window is wide enough and falls back to a floating panel when it is not; selecting text raises a small dialog that sends your question along with the selection, its surrounding context and the abstract. "Create project" materialises the paper into a workspace as paper.pdf, paper.md, paper.html, the figures and a README. Every request its browser half makes goes to its own routes on your host — there is no outbound call anywhere in the build. What you must supply is poppler: pdftotext, pdftoppm and pdfimages have to be on PATH (brew install poppler, or apt-get install poppler-utils), and without them nothing converts. Published 2026-08-15 as one version. Maintained by louwenbo580, not by us.

Four tools. cite_lookup resolves a DOI to its record; cite_format renders a reference in GB/T 7714 — the Chinese national standard most domestic journals require — as well as APA, MLA and Chicago; cite_bibtex emits BibTeX; cite_check reads a manuscript and reports citations that do not match the list, entries the text never cites, and DOIs that do not resolve. It calls Crossref and nothing else, needs no API key, and ships no runtime dependencies. Maintained by STARDUSTLC666, not by us.

Five tools that climb from discovery to evidence. zotero_search finds items by title, creator, year or indexed full text; zotero_get reads one item's metadata with previews of its notes, annotations and attachments; zotero_retrieve returns the passages that answer a question, drawn from your own annotations and notes as well as the abstract and full text; zotero_attachment resolves an item to the PDF's verified path on disk; zotero_export hands the work to Zotero's own citation engine for APA/CSL bibliographies, BibTeX, biblatex, RIS or CSL-JSON. It talks to exactly one address — Zotero's local API at 127.0.0.1:23119 — needs no API key, and has no path that writes to your library. You must enable Settings → Advanced → "Allow other applications on this computer to communicate with Zotero" in Zotero 7 or later, otherwise the tools have nothing to reach. Maintained by Vncntvx, not by us.

Four tools for the revision loop on a LaTeX manuscript: paperlab_read_state reads where the paper stands, paperlab_apply_edits writes changes back, paperlab_compile_check builds it and reports what broke, and paperlab_finish_revision closes the round. The intended shape is that you annotate the PDF and the agent works through your marks. Zero runtime dependencies, and the only address in the code is 127.0.0.1. Maintained by maple-pwn, not by us.

Four tools that read a draft and report on it rather than rewriting it: writing_audit finds leftovers from editing and the tells of generated text, writing_rules states the rules being applied, and the two profile tools let you set a house style and keep a record across drafts. Bilingual, zero runtime dependencies, and no network calls — your unpublished manuscript never leaves the machine, which for this kind of tool is the property that matters. Maintained by xmutfyh, not by us.

A cot-trim skill and a zero-dependency checker for one specific embarrassment: the reasoning, asides and half-applied edits that survive a copy-paste from an AI session into a manuscript. It reads files and reports; it makes no network calls and needs no key. Narrower than dsh-plugin-writing-guard and worth having alongside it rather than instead of it. Maintained by YuanyuanMa03, not by us.

Two halves in one row. The browser half draws a small toolbar when the pointer rests on a rendered formula; the host half serves MathJax's ~2.5 MB es5 tex-to-svg bundle from inside the package under /dsh-latex-tools/mathjax, fetched lazily on the first export so the always-loaded client stays small, then cached by the browser. The asset route rejects path traversal before it reads anything, and the only URLs anywhere in the build are the two SVG namespaces — it makes no request of its own. Its single dependency is mathjax itself. Small and finished rather than ambitious: it does one thing that is otherwise genuinely annoying, which is getting a formula out of a chat window and into your paper. Maintained by liuup, not by us.

Adds a Research plugins page to the Settings sidebar. Every install is confirmed first — the dialog names the plugin, the profile it lands in, and the exact command — and only packages this catalog lists can ever be spawned; the browser half may ask for any string, the host checks it against the allowlist. Installs pin the catalog version (name@version), the one whose code was read for the listing. From 0.4.0 the panel compares each installed package’s on-disk version with the catalog and offers an update button, itself included. The panel makes exactly one network request: fetching this catalog. Maintained by AI4Scholar.

A notebook-style bioinformatics workbench: nine bio_* tools that create projects, add and run cells, rerun one after an edit, delete a cell, record feedback and list what exists. It ships a browser half, so the workbench appears in the interface rather than only in tool calls, and a bio-workbench skill that tells the agent how to drive it. Zero runtime dependencies; the only address in the code is localhost. Maintained by poplarity, not by us.