ARIS ⚔️🌙 Auto-Research-In-Sleep
Lightweight Markdown-only skills for autonomous ML research: cross-model review loops, idea discovery, and experiment automation.
Evidence shows the repository contains multiple components (e.g., llm-chat, feishu-bridge) that require API keys and network access, but least privilege or user confirmation mechanisms are not clearly stated. Data flow transparency is limited; sensitive data handling (e.g., API keys) is mentioned but not detailed. Dependency security is not fully assessed; external effects (e.g., network calls) exist but are not explicitly controlled. Rollback mechanisms are not mentioned. Source attribution is partially clear (e.g., MIT license, author info) but unverified.
Internal consistency is good; README and code structure are largely aligned, but dependency availability is not explicit (e.g., requires external API keys). Failure messages have some examples (e.g., error handling) but are not comprehensive.
Target audience is clear (AI researchers, developers), scenarios are diverse (research automation, interview prep, etc.). Capability boundaries are partially described (e.g., skill list), but trigger precision (e.g., skill trigger conditions) is not detailed. Environment fit is good (supports multiple CLIs and OS).
Information architecture is clear (README, docs, skills directories), install notes exist (e.g., ARIS-Code CLI installation), naming stability is moderate (frequent version updates), examples and FAQ are provided, known limitations are partially mentioned (e.g., output folding), license is clear (MIT), versioning/changelog is detailed, maintenance responsibility is partially clear (author and contributors).
Output usability is good (e.g., generates reports, HTML), marginal value is high (automates research workflow), cost-benefit is not fully assessed (e.g., API costs).
Claim traceability is moderate (e.g., technical report link), cross-source corroboration is limited (e.g., external recommendations), fact-inference separation is unclear.
- The repository contains components that require API keys; ensure secure storage to avoid leakage.
- Dependencies on external services (e.g., OpenAI, Feishu) may incur network and cost implications.
- Frequent version updates; monitor changelog for behavior changes.
What does this agent do, and when should you use it?
ARIS is a lightweight, Markdown-only skillset for autonomous ML research. It orchestrates cross-model collaboration: Claude Code drives the research while an external LLM (via Codex MCP) acts as a critical reviewer. Its core methodology is an adversarial review loop to break single-model blind spots. The skill layer is transportable across Claude Code, Codex CLI, Cursor, Trae, Antigravity, and GitHub Copilot CLI. It is also packaged as a standalone ARIS-Code CLI (Rust) providing a REPL, MCP tool dispatch, no-API-key cross-model reviewing, and a bundled skillset. The repo includes 81 skills, 13 named workflows, and extension projects like Anti-Autoresearch integrity forensics and ARIS-Movie-Director. Licensed MIT.
It provides a collection of skills as plain Markdown, each defining a step of autonomous research: idea generation (/research-idea), literature review (/research-lit), paper writing (/paper-writing), experiment design (/experiment-bridge), and post-acceptance tasks like slides generation (/paper-slides) and rebuttal (/rebuttal). The core mechanism is cross-model review: Claude Code executes, while another model (via Codex MCP or LLM API) audits the output. Skills are composable into pipelines, such as /research-pipeline, which takes a research direction, reads reference papers, clones repos, generates ideas, runs experiments, and writes the paper. The standalone ARIS-Code CLI offers a REPL (/plan mode, /tasks persistent tracking), tool calls, MCP server integration, and auto-setup with bundled skills. It also includes a Research Wiki for persistent note-taking across sessions.
- An ML researcher who wants automated research idea generation, experiment running, and paper drafting overnight, waking up to a scored paper.
- A graduate student with a paper draft and reviews who needs a structured rebuttal within character limits with fabrication safeguards.
- A researcher with a codebase and an idea who wants to automate the full improvement pipeline using /research-pipeline.
- A team that wants to use cross-model adversarial review to avoid single-model blind spots, with Claude + Codex or alternative model pairs.
- A developer who wants to bring the same skill workflows to non-Claude environments: Codex CLI, Cursor, Trae, or other SKILL.md-compatible agents.
- An educator or author who wants to collaboratively write teaching materials or structured documents with cross-review using /render-html.
What are this agent's strengths and limitations?
- Cross-model review breaks self-play blind spots: two models (speed × rigor) yield better outcomes than a single model talking to itself
- Fully transportable: plain Markdown SKILL.md files can move between Claude Code, Codex CLI, Cursor, Trae, Copilot CLI, etc., with zero lock-in
- Built-in rebuttal safety gates: /rebuttal has no-fabrication, no-overpromise, and full-coverage guarantees
- No-API-key cross-model reviewing: via Codex MCP with ChatGPT subscription, you can have a reviewer without an OpenAI API key (or use alternative models like Kimi, DeepSeek)
- Active development with transparency: detailed changelogs referencing real user issues, and frequent releases
- Steep learning curve: multiple workflows, skills, and configuration options may overwhelm beginners
- Relies on external model APIs: ensuring reliability requires API credentials or multiple providers; cost can vary with plans
- Potential for hallucinations like any AI tool: despite safeguards, outputs need human verification, especially for critical research
- Requires a compatible agent runtime: although not locked in, it needs a SKILL.md-compatible agent like Claude Code or Codex CLI
How do you install or deploy this agent?
There are several installation paths:
- Lightweight skills: clone the repo and run
bash tools/install_skills.sh(or the relevant installer for your agent, e.g.,install_skills_codex.sh). The installer copies skills to~/.claude/skillsor~/.config/codex/skills. - Standalone CLI: download the latest ARIS-Code binary from the releases page, or build from source (requires Rust). Run
aris setupto configure provider and model. - Via script:
bash tools/smart_update.sh --applyto pull updates. For selective install, use--groups X,Yor--skills Xflags.
How do you use this agent?
For skills: start in Claude Code, invoke /research-pipeline "research direction" for the full pipeline, or individual skills like /research-idea, /research-lit, /paper-writing. Cross-model review requires configuring a second model: via aris setup (option 10: Codex MCP uses ChatGPT subscription, no API key) or with OpenAI/Anthropic API keys. For the ARIS-Code CLI: run aris "your prompt" or aris to start a REPL. Use /plan for planning, /tasks for tracking tasks, /research-wiki for persistent memory.
How does this agent compare with similar options?
It is compared to Anthropic's Claude Science skills and OpenAI's Codex CLI native skills. It encourages use with Codex CLI over Claude Code for greater model flexibility. It distinguishes itself from commercial tools like 'ChatGPT autopilot research' by emphasizing its lock-in-free approach.