yzlin286-wq/Deepscient

A Chinese-language scientific research workbench running on the DeepSeek Harness (dsh) runtime: it structures topic selection, retrieval, experimentation, analysis, plotting, writing, submission, and reviewer response into 12 stages with 55 decidable gates, equipped with 36 research tools and 98 skills, emphasizing evidence chains and reproducibility.

A Chinese-language scientific research workbench running on the DeepSeek Harness (dsh) runtime: it structures topic selection, retrieval, experimentation, analysis, plotting, writing, submission, and reviewer response into 12 stages with 55 decidable gates, equipped with 36 research tools and 98 skills, emphasizing evidence chains and reproducibility.

Stars 335 (+96 / 24h)

Reasons

  • Each gate is evaluated against the actual state in the repository, rather than relying on the model's subjective judgment; experiment records, environments, versions, and random seeds are persisted to disk, allowing reruns on different machines.
  • Built-in literature sources include Crossref, OpenAlex, arXiv, PubMed, Semantic Scholar, and Europe PMC; an optional MCP layer adds multi-source retrieval, downloading, and PubMed full-text tools; additionally, it covers journal quartiles, cycles, and submission requirements for over 5000 journals.
  • For aspects the system cannot compute (construct validity, novelty, ethical trade-offs), it provides checklists without drawing conclusions for you; the interface features five tabs: Workbench, Project Library, Integration, Billing, and Documentation.

Usage

  • For Windows, download DeepScient-desktop-v0.12.0.zip (approx. 635 MB), extract it to a short path, and double-click "DeepScient Start.cmd"; for Apple Silicon Macs, use the dmg for 0.12.0.
  • The source code method requires Windows 10/11, Node 22.5+, and an installed DeepSeek Harness desktop version: install dsh-science within dsh, run node scripts/deepscient-sync-models.mjs, then launch deepscient\bin\deepscient.cmd, and open 127.0.0.1:3081 in your browser.
  • Third-party research skills must be installed separately: run node deepscient/scripts/install-skills.mjs --plan to view the list first, then use --apply.

Risks

  • Depends on the DeepSeek Harness runtime and your configured model providers; initial setup requires browser login and authorization; the macOS version is unsigned and requires manual approval.
  • Of the 98 skills, 81 come from third-party repositories (the author states they are pinned to specific commits; see docs/third-party.md for sources and licenses); the repository was created on 10 21, under the MIT license.

Open

GitHub

Explore the latest AI Radar updates