Auto Claude Code Research in Sleep

ARIS (Auto-Research-In-Sleep) is a set of lightweight, markdown-only skills that turn Claude Code, Codex CLI, Cursor, and similar coding agents into an autonomous research loop. One model acts as the executor, driving the work; a second, independent model reviews its output at each step. It's built for ML researchers, engineers running agentic coding tools, and anyone who wants an agent to run unattended research or implementation cycles without silently drifting off course.

The project describes itself as a methodology rather than a platform: the skills are plain markdown files you drop into whatever harness you already use. There's no proprietary runtime lock-in. The same workflow ships as skills for Claude Code, Codex CLI, Cursor, Trae, Antigravity, GitHub Copilot CLI, OpenClaw, and a DeepSeek Harness variant, plus a standalone CLI (ARIS-Code) and Claude Code / Codex CLI plugins for one-command setup.

Key features

• Cross-model review loops: the executor model proposes work, and an independent reviewer model (Codex MCP by default) checks it before it's accepted, catching drift and unjustified shortcuts.
• Markdown-only skills: the entire workflow is defined in plain markdown skill files, so it's portable across agent harnesses without custom tooling.
• Harness-agnostic: works inside Claude Code, Codex CLI, Cursor, Trae, Antigravity, Copilot CLI, OpenClaw, and DeepSeek Harness, each with its own adaptation docs.
• Alternative model combinations: supports model pairings that don't require Claude or OpenAI API keys, including ModelScope-hosted models.
• Standalone CLI (ARIS-Code): a packaged CLI with session resume, output folding, reasoning-effort controls, and model aliasing, distributed via GitHub Releases.
• Plugin distribution: installable as a Claude Code plugin or Codex CLI plugin for a faster setup path than manually copying skill files.
• Companion tooling: a small macOS menu-bar widget (ARIS-Monitor) flags when a session is waiting on your approval.
• Sibling projects: related repos apply the same executor+reviewer methodology to interview prep content, over-defense mitigation in agent prompts, and multimodal storyboard generation.

Ideal use cases

ARIS fits teams running long, multi-step ML research or engineering tasks through an agentic coding tool and who want a built-in check on the agent's output instead of trusting a single model's self-review. It's a good match if you already use Claude Code, Codex CLI, Cursor, or one of the other supported harnesses and want to add a review discipline without switching tools.

It also suits people experimenting with idea generation or paper-to-code workflows who want the agent to iterate overnight ('in sleep') while a second model gates each step.

It is not a good fit if you need a hosted platform with a dashboard and no agent setup; ARIS requires you to wire the skills into an existing coding agent yourself. It's also not useful if you only need a single quick script run, since the value comes from the review loop over longer, multi-step work. Teams without API access to at least two models (or a supported no-API-key model combination) will have a harder time using the reviewer step as intended.

Installation

ARIS is distributed three ways: as markdown skill files you drop into your agent's skills directory, as a plugin for Claude Code or Codex CLI, or as a standalone CLI.

To use the skills directly, clone the repository and copy the relevant skill folder into your harness's skill directory for your tool (Claude Code, Codex CLI, Cursor, Trae, Antigravity, Copilot CLI, OpenClaw, or DeepSeek Harness each have their own adaptation docs in the repo):

git clone https://github.com/wanshuiyin/Auto-claude-code-research-in-sleep.git
cd Auto-claude-code-research-in-sleep

For the one-command setup path, install the Claude Code plugin or Codex CLI plugin referenced in the repo's plugin section.

For the packaged CLI, download the latest ARIS-Code binary from the project's GitHub Releases page and follow the CLI README for your platform.

By default the reviewer step uses Codex MCP, which needs its own API access configured; the repo documents alternative model combinations, including ModelScope-hosted models, for setups that avoid requiring Claude or OpenAI API keys.