Summary

Luxas is an open-source, multi-agent system designed for autonomous scientific research. It automates the entire process from a research question to a compiled LaTeX manuscript, including literature surveys, experiment design, execution, and report generation with figures and citations. Aimed at researchers and power users, it provides a robust, unattended workflow that can run for hours, featuring crash recovery and adversarial review.

How It Works

Luxas functions as a harness built on pi-mono primitives, orchestrating specialized agents (e.g., search, reader, experiment, illustrator) via a central "brain" agent. It employs file-backed memory, detached Node sub-processes, and deterministic finish-gates. Its approach is novel in its robust crash-recovery (replaying from logs), multi-model LLM support (Anthropic, DeepSeek, Kimi, OpenAI), and adversarial review across content, figures, and layout, ensuring a high-quality, reproducible PDF output.

Quick Start & Requirements

• Primary install: npm install && npm link or npx tsx src/index.ts.
• Prerequisites: Node.js 22+, API keys (Anthropic default; DeepSeek, Kimi, OpenAI, Wolfram, Brave, Gemini optional), LaTeX (mactex or texlive-latex-extra), poppler, Python 3.10+ (with matplotlib, numpy), tmux. Optional: wolframscript, provref, browser-use for paywalled venues.
• Demo: Try in browser at luxas.im; example reports at luxas.im/gallery.

Highlighted Details

• Automated research pipeline: literature survey, experiment design/implementation/review, and LaTeX report generation with figures and citations.
• Crash-recoverable harness: Replays from logs, detached sub-agents, and orphan recovery enable unattended, multi-hour runs.
• Flexible LLM integration: Supports Anthropic, DeepSeek, Kimi, and OpenAI models, configurable via environment variables for cost/performance optimization.
• Rigorous adversarial review: Three layers of review (content, figure internals, PDF layout) ensure quality and catch subtle regressions.
• Reproducible artifacts: Generates compiled LaTeX PDFs with number-provenance for auditable results.

Maintenance & Community

Developed primarily by Mu Qiao (Muuuun), with contributions noted from Mario Zechner for pi-mono. Receives token sponsorship from Deeplang 深言科技. No explicit community channels (e.g., Discord, Slack) are listed in the README.

Licensing & Compatibility

• License: MIT.
• Compatibility: The MIT license is permissive, generally allowing commercial use and integration with closed-source projects without significant restrictions.

Limitations & Caveats

Luxas is a specialized tool for research topics requiring literature surveys and computational studies culminating in reports, not a general-purpose agent framework. Setup requires significant external dependencies and API key configuration. The default LLM profile can incur $20-80 per run, though cheaper alternatives exist. Project directories should be treated as executable code, as it is not a sandboxed environment.