skyllwt/AutoSci

What it solves

AutoSci is designed to automate the human-intensive process of scientific research. It handles the full research lifecycle, including literature ingestion, idea generation, experimental design, execution, and manuscript writing, while maintaining a persistent, structured memory that evolves across different projects.

How it works

AutoSci operates as a memory-centric agentic system that integrates with coding-agent runtimes like Claude Code, Codex, or OpenCode. It uses a set of specialized "skills" (commands) to manage the research process:

• Knowledge Management: It ingests papers and notes to build a research wiki and knowledge graph, which can be visualized via a Web UI or Obsidian.
• Ideation and Experimentation: It uses commands like /ideate to generate research directions and /exp-design, /exp-run, and /exp-eval to manage the end-to-end experimental pipeline (from pilot runs to final evaluation).
• Discovery: It features /discover to find relevant papers from specific venues (e.g., ICLR, NeurIPS) and /daily-arxiv for scheduled paper recommendations based on the user's wiki interests.
• Publication: It can automatically transform a LaTeX paper draft into a conference poster using the /poster skill.
• Review System: It supports a cross-model review process where a second LLM acts as an independent reviewer for ideas and drafts.

Who it’s for

It is intended for researchers and scientists who want to automate the repetitive and coordination-heavy parts of the research lifecycle, from literature review to paper writing.

Highlights

• Full Lifecycle Automation: Covers everything from initial paper ingestion to the final rebuttal.
• Persistent Research Memory: Maintains a structured memory that compounds across projects.
• ** uma-centric Workflow**: Integrates with professional coding agents (Claude Code, Codex, OpenCode) to execute code and run experiments.
• Automated Poster Generation: Converts LaTeX sources into print-ready HTML/PNG posters.
• Knowledge Graph Visualization: Provides both a browser-based UI and Obsidian integration for exploring research connections.