What it does ARIS (Auto-Research-In-Sleep) is a set of lightweight, Markdown-based skills that plug into existing LLM agent runtimes — Claude Code, Codex CLI, Cursor, Trae, even GitHub Copilot CLI. It orchestrates autonomous ML research workflows: idea discovery, paper review, experiment automation, and persistent research wikis. There’s no framework to adopt and no vendor lock-in; the same workflow files run across different agents.

The interesting bit The project treats “research” as a structured adversarial process. A breadth pass (parallel fan-out or agent spawn) generates candidates, a cross-model jury reviews for accuracy, and a research wiki serves as memory. The README’s own release notes are almost comically detailed — v0.4.15 fixed SSE EOF handling because MiniMax doesn’t send data: [DONE], and v0.4.14 redacted system prompts that were leaking Bearer tokens. That level of operational paranoia suggests someone is actually running this at scale.

Key highlights

• Multi-runtime: Skills for Claude Code ultracode, Codex spawn_agent, Cursor, Trae, OpenClaw, and standalone CLI
• Cross-model review: Loops report to the same jury + wiki regardless of which agent drives them
• Standalone CLI: ARIS-Code with 11 releases since May 2026, adding DeepSeek V4 Pro, Qwen 3.6, Xiaomi MiMo, and custom OpenAI-compatible providers
• Security-hardened: System prompt secret redaction, sandbox strict mode, URL origin stripping for MCP servers
• Ecosystem sprawl: Spin-offs for non-academic research (ARIS-Anything), AI interview prep (ARIS-in-AI-Offer), and even a macOS floating monitor widget (ARIS-Monitor)

Caveats

• The README is dense with Chinese-English hybrid prose, future-dated references (“VALSE 2026”), and a sprawling ecosystem that can be hard to navigate
• MCP tool dispatch is still experimental as of v0.4.15; the aris doctor command warns when MCP servers are configured
• The “no framework” claim is technically true, but the skill system has grown enough moving parts that it approaches framework territory

Verdict Worth a look if you’re already living in Claude Code or Codex and want structured research automation without switching tools. Skip it if you want a polished SaaS — this is a power-user tinkerer’s toolkit with sharp edges and a 77-skill bundle to maintain.