FAROS

FAROS (Foundation AutoResearch Operating System) is an open-source multi-agent system from OpenNSWM-Lab that covers research topic selection, literature retrieval, planning, code generation, experiment tracking, paper writing and peer review. It keeps provenance, decisions, artifacts and human feedback at each stage.

What It Is

FAROS is a research-workflow runtime and web workbench for running an AI-assisted research loop. A user starts from a research interest, and the system moves through Idea, PlanPackage, Code, Experiment, Paper and ReviewX stages, ending in human approval or an evidence-driven revision. Its runtime uses a Blueprint + Capability + Profile + Provider design so new workflows and tools can be added.

How the research loop works

• Idea: topic coaching turns a rough interest into a task, method, dataset and evaluation target; literature retrieval, semantic filtering and deep reading run before ideation.
• PlanPackage: a structured contract carrying hypotheses, variables, steps, acceptance criteria and evidence references, with reviewer checks and human approval.
• Code and Experiment: project generation, sandboxed execution, static and dynamic evaluation, metric ingestion, dataset upload and publication-style figures.
• Paper: brief, outline, section drafting, citation management and LaTeX/PDF generation linked to plans and experiments.
• ReviewX: extracts claims, aligns them with evidence, checks consistency and reliability, and produces revision actions.

Setup path

The backend is a FastAPI app (Python 3.11+) and the frontend is a React/TypeScript app (Node.js 18+). Models are configurable per user through provider settings or environment variables; the README lists OpenAI, Anthropic, DeepSeek, Moonshot, Zhipu and MiniMax adapters and compatible endpoints. Docker Engine is required for the formal Code/Experiment sandbox, and formal PDF output needs latexmk, XeLaTeX, ctex and a CJK font. The README also describes user-scoped provider settings with encrypted API key storage, a bilingual UI and recoverable background jobs.

License and deployment

The current repository licenses original source and documentation under MIT. Setup instructions target Linux or WSL Bash, with Python 3.11+ and Node.js 18+. Model/API and compute costs are separate. Important scientific conclusions still require human verification against their evidence.