Luxas Research Colleague
Turns a research question into a cited, illustrated, compiled LaTeX manuscript through a file-backed multi-agent workflow.
Per-dimension scores and reasoning
Evidence shows: README details safety mechanisms including file write protection, cost limits, finish gates, sub-agent isolation, and blocking access to sensitive files (e.g., ~/.ssh/id_*), reflecting least privilege. However, user confirmation is weak, only via --max-cost, no explicit approval steps. Data flow transparency is good, README explains state storage and logging. Sensitive data handling has guards but no encryption mentioned. Dependency security is insufficient, versions pinned but no vulnerability scanning mentioned. External effects are controlled via cost limits and finish gates. Rollback is limited, only crash recovery, no explicit rollback. Source attribution is clear, README and LICENSE identify author.
Evidence shows: README describes internal consistency via finish gates and plan commitments. Dependency availability is at risk due to external APIs and models, no fallback mentioned. Failure messages: README mentions crash recovery and error handling but no concrete examples.
Evidence shows: README identifies target audience (researchers) and scenarios (literature survey, experiment design). Capability boundaries are clear, stating when to use and not use. Trigger precision is high via RESEARCH.md and CLI args. Environment fit is good, with install instructions for macOS and Linux.
Evidence shows: Information architecture is clear, README well-structured. Install notes are detailed. Naming stability is moderate, version 0.1.0 but no API stability mentioned. Examples and FAQ are rich. Known limitations are mentioned (e.g., cost, model dependency). License is MIT. Versioning/changelog missing. Maintenance responsibility is with author but not explicit.
Evidence shows: Output usability is high, generates compiled PDF reports. Marginal value is high, automates research workflow. Cost-benefit is explained with cost estimates.
Evidence shows: Claim traceability is good, README mentions provref for number provenance. Cross-source corroboration is insufficient, no external verification. Fact-inference separation is good, README distinguishes observations and inferences.
- Reliance on external APIs and models poses vendor lock-in and cost risks.
- Security mechanisms are defense-in-depth, not a sandbox, and cannot fully prevent malicious code execution.
- User confirmation is weak; autonomous runs may have unintended consequences.
What does this agent do, and when should you use it?
Luxas is an open-source multi-agent system for taking a topic in RESEARCH.md through literature work, experiments, and a compiled LaTeX PDF. Its brain coordinates detached Node sub-agent processes for search, reading, experiments, mathematics, illustration, typesetting, and adversarial review. Project state is externalized in files, logs, and notes; a resumed run reconstructs accounting from log.jsonl and recovers background work. The intended artifact is a cited report with self-generated figures and review notes, suited to literature surveys and small computational studies. It runs as a self-hosted CLI workflow in a local project directory, with shell, network, LaTeX, Python, and provider credentials rather than as a hosted chat application.
After luxas init creates a project, brain decomposes the research brief. search discovers papers through OpenAlex, arXiv, CrossRef, citation chains, web search, and an optional anti-detect browser; reader writes per-paper material to notes/literature.d/ for consolidation into notes/literature.md. experiment designs tools, then assigns scripts/<tool>.py to tool_impl and tests/test_<tool>.py to tool_review independently; pytest failures are sent back for revision before results are written to data/experiments/<EXP_ID>/runs/run_N/results.json and notes/experiments.md. math can call Wolfram Engine through wolframscript and falls back to sympy, while illustrator, illustrator_write, and typesetter generate or audit figures and rasterized PDF pages. The finish tool checks completed experiment commitments, background agents, report.pdf, a self-generated figure, typesetter clearance, and PI review status before allowing a clean completion.
- A researcher who needs a cited survey PDF from one well-scoped scientific question.
- A computational research group that wants experiment tooling designed, independently implemented and tested, then incorporated into a manuscript-style report.
- An individual running multi-hour unattended research jobs who needs file-backed logs, notes, and restart recovery.
- A research engineer who wants independent implementation and test authors for experiment scripts rather than one agent validating its own semantics.
- A technical user with local LaTeX, Python, tmux, and API credentials who wants to run research automation in a project directory.
What are this agent's strengths and limitations?
- Externalized project state, reverse-scanned log recovery, detached processes, and orphan recovery support long-running work that can be resumed after a crash.
- The experiment workflow separates tool_impl from tool_review and uses pytest as the validation ground truth.
- It joins literature discovery, experiment execution, figure work, LaTeX compilation, figure review, and page-layout review in one research-oriented delivery path.
- Aligned finish and reviewer gates require completed experiment records, artifact checks, and review conditions instead of treating a model's assertion of completion as sufficient.
- Adoption requires a local shell, filesystem workspace, Node.js, LaTeX, Poppler, Python with plotting packages, tmux, and API credentials; it is not a zero-install hosted service.
- Anthropic is the default provider. DeepSeek and Kimi alternatives have configuration and capability tradeoffs: the documented DeepSeek path is text-only, so vision work needs a separate vision profile.
- The system may autonomously run Python, shell commands, and pip install. The documentation advises against targeting directories containing credentials and against running as root.
- Documented run costs are anecdotal and variable: about $20–80 with the default Claude setup and $2–10 with the dual profile, depending on topic depth and review iterations.
How do you install or deploy this agent?
Install the documented prerequisites first. On macOS:
brew install --cask mactex
brew install poppler tmux [email protected]
pip3 install matplotlib numpyThen install Luxas:
git clone https://github.com/Muuuun/luxas.git && cd luxas
npm install && npm linkNode.js 22+ and ANTHROPIC_API_KEY are required for the default path. pdflatex, bibtex, Poppler, Python, matplotlib, numpy, and tmux must be available. OPENAI_API_KEY is optional for the o3 math agent; DeepSeek, Kimi, Brave, Gemini, Wolfram Engine, browser-use, and provref are optional according to the features used.
How do you use this agent?
Set the default provider key, initialize a project, and start a run:
export ANTHROPIC_API_KEY="..."
luxas init ~/research/x --prompt "Survey LLM chain-of-thought reasoning"
luxas run ~/research/x --model opus
luxas status ~/research/xUse "luxas figures ~/research/x" to rerun only the figure/typesetter loop and "luxas list" to list projects. The default uses each agent definition's Claude model; "luxas run ~/research/x --profile dual" selects deepseek-v4-pro for text and k2p5 for vision-required work.
How does this agent compare with similar options?
Luxas is positioned as a research-specific system rather than a general workflow framework. Against LangGraph, CrewAI, and AutoGPT, it emphasizes file-backed state, hook-enforced closure gates, and a compiled LaTeX PDF deliverable; against Sakana AI Scientist, its documentation emphasizes literature discovery and citations; against Claude Code, it uses multiple detached agent roles instead of one chat session. For general-purpose orchestration or interactive coding, the repository points readers toward LangGraph/pi-agent-core or Claude Code respectively.