ARIS ⚔️ (Auto-Research-In-Sleep) is a lightweight solution for autonomous machine learning research, whose main body consists of skills (Skills) written in pure Markdown. According to the project description, ARIS emphasizes that "this is a methodology, not a platform," with the focus on the research workflow itself, so users can bring it to different tools. It does not depend on a specific framework and does not create platform lock-in, and can be used with LLM agents such as Claude Code, Codex CLI, Cursor, Trae, Antigravity, GitHub Copilot CLI, OpenClaw, DeepSeek Harness, and others.
In terms of purpose, ARIS targets machine learning research scenarios, covering cross-model review loops, research idea discovery, and experiment automation. Its basic mode of operation is: an executor advances the research work, and then an independent model reviews it (the default reviewer is Codex MCP). The project mentions that alternative model combinations can be used, including models hosted by ModelScope, so it can be used without Claude or OpenAI APIs.
In terms of usage, ARIS provides two paths: first, installing the skills directly into the above agent environments for use; second, one-click installation and use through a standalone CLI or plugin. The repository also includes a standalone installation version of the ARIS-Code CLI, and provides a download entry and Chinese and English documentation. The CLI has a version update history involving capabilities such as the review bridge, session management, model switching, and proxy support. For example, v0.4.25 states that codex-cli 0.154.0 removed codex mcp-server, and ARIS instead calls through a self-built bridge on top of codex exec, with the number of skills and results remaining consistent, and prompts users to re-register according to the Quick Start steps.
Applicable scenarios include: needing to continue advancing ML research tasks unattended or during "sleep" periods; needing different models to cross-review research conclusions and code; needing to string idea discovery, experiment execution, and result review into a reusable process; and migrating the same research workflow between multiple agent hosts.
In terms of limitations: the materials do not provide specific pricing plans, price details, or commercial terms; ARIS itself is described as a lightweight solution whose core is Markdown skills rather than a complete platform, so its capability boundaries depend on the connected agent and model. The registration steps, model combinations, version changes, and other items mentioned in the materials are the project's own installation and maintenance requirements, and specific availability may change with the host tool and dependency versions. For metrics such as evaluation data, supported languages, concurrency scale, and accuracy, the materials do not currently provide them.
