Auto-claude-code-research-in-sleep
ARIS ⚔️ (Auto-Research-In-Sleep): lightweight Markdown-only skills for autonomous ML research: cross-model review loops, idea discovery, and experiment automation. No framework, no lock-in. Works with Claude Code, Codex, OpenClaw, or any LLM agent.
The Lens
By Erik Loyd, SaaS CEO and former COO/CFO of an AWS Premier Partner.
Updated Oct 2026
ARIS (Auto-Research-In-Sleep) is a free MIT pack of Markdown skills that turns Claude Code into an overnight ML research assistant. Give it a direction, or a paper plus code, and it finds ideas, runs experiments, writes the paper and drafts rebuttals. The twist: Claude does the work and a different model, GPT via the Codex CLI by default, reviews it, since a model grading itself misses its blind spots.
Setup: clone the repo, link 80-plus skills into your project, register the Codex reviewer over MCP. Papers need LaTeX; experiments need a GPU over SSH, a local card, Vast.ai or Modal. It also runs in Codex, Cursor or its own CLI.
No paid tier, but models cost money. By default it rides your Claude Code and ChatGPT logins, so a night burns plan quota. On API keys, Claude Opus 5 is $5 per million tokens in and $25 out, and the review loop runs up to four rounds by default, plus GPU time. Solo researchers and small labs: worth it. Larger teams: no shared service.
The catch: overnight runs mean pre-approving file writes, and the loop chases a reviewer's score, not the truth. Check every claim and citation before submitting.
Free vs Self-Hosted vs Paid
fully freeFree: All of it under MIT: 84 skill folders covering idea discovery, literature search, experiment planning and execution, the auto review loop, paper writing, rebuttals, conference talks, patents and Lean proofs, plus a persistent research wiki, a Codex reviewer bridge (mcp-servers/codex-exec), and the standalone ARIS-Code CLI (MIT, Rust, prebuilt for macOS, Linux and Windows).
Self-hosted: Runs wherever your agent runs; there's no server. Clone the repo, run the install script to link skills into .claude/skills, and register the reviewer with claude mcp add codex. Paper writing needs LaTeX (latexmk, pdfinfo). Experiments need a GPU: remote SSH, local, Vast.ai on demand, or Modal. Overnight runs also mean pre-approving file writes and reviewer calls in .claude/settings.local.json.
Paid: No ARIS tier. You pay for models and compute. Default route: Claude Code as executor plus GPT reviewing via codex login with a ChatGPT account, so it consumes subscription quota rather than per-token billing. On API keys, Claude Opus 5 is $5/$25 per million input/output tokens (Opus 5.5: $4/$20). The auto review loop runs up to 4 rounds by default (MAX_ROUNDS), each a deep xhigh-effort GPT review plus Claude rewriting and possibly launching experiments. The README says experiments over 4 GPU-hours are skipped, but the current auto-review-loop skill file doesn't encode that cap, so don't count on it. Cheapest route: ModelScope's free DeepSeek-V3.1 (2,000 calls/day) or effort: lite.
Free under MIT. A night of runs costs whatever your models and GPUs cost: plan quota on Claude and ChatGPT subscriptions, or per-token bills on Opus-class API keys, plus GPU time.
What to do by team size
- Solo
- free; run it on the Claude Code and ChatGPT plans you already pay for
- Small team
- free; each researcher runs their own setup and model budget
- Medium team
- free, but model spend and GPU time scale per person with no central controls
- Large team
- free, though there's no shared service, admin or audit trail; treat it as a per-researcher tool
Get tools like this every Wednesday
One featured tool, three on the radar. No fluff.
Similar Tools
Claude Autoresearch Skill: autonomous goal-directed iteration for Claude Code. Inspired by Karpathy's autoresearch. Modify → Verify → Keep/Discard → Repeat forever.
Fully autonomous & self-evolving research from idea to paper. Chat an Idea. Get a Paper. 🦞
Autonomous experiment loop extension for pi
Structured prompts, vault templates, and archive guides for AI-assisted genealogy research. Built for Claude Code.
A low score is not a verdict on quality. Young and niche tools start low by design. How we calculate scores
Trust Signals
License: MIT License
Use freely, including commercial. Just keep the license.
Commercial use: ✓ Yes
About
- Owner
- Yang Ruofeng (User)
- Stars
- 18,759
- Forks
- 1,523
- discussed