Scrinium

A research infrastructure for AI agents.

English | 中文

"

Fork: Scrinium is a hard fork of ScholarAIO (MIT). Original work © Zi-Mo Liao — see LICENSE.

Your coding agent already reads code, writes code, and runs experiments. Scrinium adds a structured research workspace on top, so the same agent can search literature, cross-check results against papers, use scientific software more accurately, and carry the whole research workflow from one terminal.

• Your paper library becomes a reusable knowledge base for the same agent.
• When scientific software questions come up, the agent can consult official documentation at runtime instead of guessing from prompts.
• The system is built to keep expanding as new tools and workflows become worth supporting.

Scrinium offers more than search. It gives an AI coding agent a research workspace that supports natural-language interaction, papers and notes, more reliable use of scientific software, writing and running code, checking results against the literature, and structured academic writing.

flowchart TB
    U["User / AI agent"] --> S["Skills .agents/skills — 35 workflows"]
    U --> C[scrinium CLI]
    S --> C
    C --> ING[Ingest pipeline]
    C --> RET[Retrieval]
    C --> CUR[Curation]
    ING --> PARSERS["MinerU / Docling / PyMuPDF"]
    ING --> APIS["Crossref / S2 / OpenAlex"]
    RET --> FTS[FTS5 keyword, BM25 field weights]
    RET --> TAGS["Curated tags + citation graph / snowball"]
    DB(["data/papers — meta.json + paper.md + notes.md"])
    ING --> DB
    RET --> DB
    CUR --> DB

Quick Start

The default and recommended way to use Scrinium is simple: install it, configure it once, and open this repository directly with your coding agent.

git clone https://github.com/wszqkzqk/scrinium.git
cd scrinium
pip install -e ".[full]"
scrinium setup

Then open the repository in Codex, Claude Code, or another supported agent. In this setup, the agent gets the fullest experience: bundled instructions, local skills, the CLI, and the complete codebase context are all available directly. For Claude Code plugins, Codex/OpenClaw skill registration, and other setup paths, see "docs/getting-started/agent-setup.md".

What It Does

Works With Your Agent

Scrinium follows the AGENTS.md open standard — open this repository in any agent that reads it and it just works. The only pointer kept is for Claude Code, which reads CLAUDE.md (a stub importing AGENTS.md) instead of AGENTS.md natively.

Skills follow the open AgentSkills.io standard and live in .agents/skills/ (with .claude/skills/ and skills/ as discovery symlinks). To reuse the skills from another project, symlink the skills directory into ~/.agents/skills/, or install the Claude Code plugin from the marketplace.

Migrating from existing tools? Import directly from Endnote (XML/RIS) and Zotero (Web API or local SQLite), with PDFs, metadata, and references brought over together. More import sources are on the roadmap.

Configuration

Start by opening scrinium with your agent and let it walk you through the setup. The notes below are only a basic overview.

Scrinium works with a minimal setup and can be expanded as needed. It makes no LLM or embedding calls of its own — all understanding comes from your agent.

• scrinium setup walks you through the basics.
• A MinerU token is optional but recommended, and free. You can also deploy MinerU or Docling locally for PDF parsing.
• scrinium setup check shows what is installed, what is optional, and what is missing.

Full setup and configuration details → "docs/getting-started/agent-setup.md", config.yaml

Agent First, CLI Available

Scrinium works best through an AI coding agent, but it also provides a CLI for scripting, debugging, and quick queries. For a current command reference aligned with the code, see "docs/guide/cli-reference.md".

Project Structure

scrinium/             # Python package — CLI and all core modules
  ingest/               #   PDF parsing + metadata extraction pipeline
  sources/              #   External source adapters (arXiv / Endnote / Zotero)

.agents/skills/         # Agent skills (AgentSkills.io format)
.agents/skills/         # ↑ symlink for cross-agent discovery
data/papers/            # Your paper library (gitignored)
data/proceedings/       # Proceedings library (gitignored)
data/inbox/             # Drop PDFs here for ingestion
data/inbox-thesis/      # Drop theses here (auto-tagged, skips DOI dedup)
data/inbox-patent/      # Drop patents here (deduplicated by publication number)
data/inbox-doc/         # Drop non-paper documents here (reports, standards, lecture notes)
data/inbox-proceedings/ # Drop proceedings volumes here for dedicated ingest

Full module reference → "docs/contributing.md"

Citation

If you use Scrinium in your research, please cite:

@software{scrinium,
  author = {Zhou, Qiankang and Liao, Zi-Mo},
  title = {Scrinium: A Research Infrastructure for AI Agents},
  year = {2026},
  url = {https://github.com/wszqkzqk/scrinium},
  license = {GPL-3.0-or-later}
}

License

GPL-3.0-or-later © 2026 Zhou Qiankang. Scrinium is a hard fork of ScholarAIO; the original work © 2026 Zi-Mo Liao remains under the MIT License.