Libra

A scientific research agent for the full research lifecycle.

From literature survey to publication-ready paper, Libra handles the mechanical parts of research so you can focus on the ideas.

---

What Libra does

A typical research workflow with Libra:

1. Survey — "Find all papers on transformer-based protein folding from 2023–2025, ranked by citation impact." Libra searches arXiv + Semantic Scholar, dedupes, and produces a BibTeX-formatted reading list with abstracts and key findings.
2. Reproduce — Drop in a paper PDF. Libra extracts the method, generates executable code, runs it in a sandboxed Docker container, and produces a reproduction report comparing your numbers against the paper's.
3. Experiment — Describe your idea. Libra designs baselines, runs experiments end-to-end, and generates publication-quality matplotlib figures.
4. Write — Hand over your results. Libra drafts a full LaTeX paper in NeurIPS / ICML / ACL templates, compiles to PDF, and manages the bibliography.

All driven by the sci-pi orchestrator subagent that chains these four skills into one continuous workflow.

Quick Start

git clone https://github.com/wisdom222/scideer
cd scideer
make config
make install
echo "DEEPSEEK_API_KEY=sk-..." >> .env
make dev

Open http://localhost:2026 in your browser.

Built On

Libra extends DeerFlow 2.0 by ByteDance — an open-source agent harness providing LangGraph orchestration, skill loader, sandbox execution, and MCP integration.

Libra adds the scientific-research domain layer:

• paper-reproduction skill — paper → executable code → comparison report
• scientific-writing skill — results → LaTeX paper → compiled PDF
• sci-pi orchestrator subagent — chains the lifecycle end-to-end
• Semantic Scholar MCP integration — citation graph beyond arXiv
• Mini sci-bench — quantitative evaluation layer for research-task performance

Repository Layout

Path	What's there
skills/	Markdown-defined research workflows
agents/	Specialized agents including sci-pi orchestrator
backend/	Python harness extending DeerFlow runtime
benchmarks/sci_eval/	Mini sci-bench evaluation suite
docs/plans/2026-05-03-scideer-design.md	Full architecture and design rationale

Acknowledgments

Libra builds on upstream work from the DeerFlow team — the harness, skill system, sandbox abstraction, and MCP client all come from there. Libra contributes the scientific-research-specific extensions on top.

License

MIT — see LICENSE.