ARIS: 16.1K stars, cross-model ML research loops for Claude Code and Codex (2026)

Auto-claude-code-research-in-sleep (ARIS) provides a structured workflow for running autonomous machine-learning research inside Claude Code and Codex. The repository bundles 82 Markdown-only skills that guide an agent through literature review, idea generation, experiment logging, and manuscript drafting. The main architectural choice is an adversarial review loop: Claude Code acts as the primary executor while OpenAI's Codex CLI reviews intermediate outputs through a Model Context Protocol (MCP) server. 1

The project counts 16.1k stars and 1.4k forks on GitHub, with active commits through September 11, 2026. The repository is released under an MIT license. The official GitHub weekly Trending snapshot for this run did not list the repository, so no verified seven-day star-growth delta is available; this review treats total stars and recent commits as signals of developer interest rather than a verified weekly breakout. 1

What the skill changes

A standard agent session relies on self-evaluation. When an agent critiques its own code or research claims, it tends to reinforce initial assumptions and miss logical oversights. ARIS addresses this by splitting the research cycle between two independent models. Claude Code executes exploratory coding, paper indexing, and text synthesis. Codex MCP inspects the resulting artifacts, challenges weak assumptions, and scores arguments before they enter persistent documentation. 1

The entire skill layer consists of plain Markdown files in a skills/ directory. Each skill operates as a standalone instruction card with zero external database or daemon dependencies. Workflows feed into an append-only knowledge base called the Research Wiki (research-wiki/), which records literature summaries, generated hypotheses, and experimental results as interconnected Markdown notes. 2

The framework supports multiple research stages:

Workflow 1 (Idea discovery): searches academic literature, maps conceptual gaps, and generates structured hypothesis cards.
Workflow 1.5 (Experiment bridge): writes training or evaluation scripts and coordinates remote GPU execution.
Workflow 2 (Auto review loop): subjects claims and experimental outcomes to adversarial review rounds.
Workflow 3 (Paper writing): converts verified claims and raw experiment tables into conference-ready LaTeX manuscripts.

Install one supported host

ARIS supports Claude Code, Codex CLI, Cursor, and related coding agents. The recommended production path pairs Claude Code as the local executor with Codex CLI as the background reviewer. 2

The setup requires two main tools:

Claude Code: installed according to official Anthropic documentation.
Codex CLI: installed and authenticated with your ChatGPT account via codex login. 2

If you plan to compile PDF papers locally, install a TeX distribution (mactex on macOS or texlive-full on Linux) along with poppler to supply pdfinfo. If your initial testing focuses on idea discovery and paper reading, you can skip the TeX installation. 2

Clone the central repository once to your home directory:

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

To attach ARIS to a specific paper or project repository, run the project installer:

cd ~/your-paper-project
bash ~/aris_repo/tools/install_aris.sh

The installer creates flat symlinks in .claude/skills/, writes an installation manifest to .aris/installed-skills.txt, and configures CLAUDE.md. The manifest ensures that uninstallation or updates only touch files managed by ARIS, leaving custom project files intact. 2

Next, register ARIS's Codex MCP bridge. Because codex-cli 0.154 removed the built-in codex mcp-server command, ARIS provides its own execution server inside the repository:

claude mcp remove codex -s user 2>/dev/null
claude mcp add codex -s user -- python3 "$HOME/aris_repo/mcp-servers/codex-exec/server.py"

Restart Claude Code immediately after running this command. Claude Code only loads MCP registrations on startup. After restarting, verify the connection in your shell:

claude mcp list | grep codex

The output should confirm codex as connected. 2

First test: verify the bridge and initialize the research wiki

Run your first test in an isolated scratch project to verify tool discovery and MCP communication without touching existing code.

Prepare a clean directory:

mkdir ~/test-aris-project && cd ~/test-aris-project
git init
touch CLAUDE.md
bash ~/aris_repo/tools/install_aris.sh

Start Claude Code inside ~/test-aris-project. First, check that the cross-model bridge functions:

Ask GPT via codex MCP: what is 1+1?

A direct response from GPT confirms that Claude Code can invoke Codex MCP for downstream reviews. 2

Next, initialize the project knowledge base:

/research-wiki init

This command creates the research-wiki/ folder structure, containing index.md, log.md, gap_map.md, and subdirectories for papers/, ideas/, and experiments/. 2

Finally, test academic paper indexing with a known reference:

/alphaxiv https://arxiv.org/abs/1706.03762

The skill fetches the Attention Is All You Need paper, extracts its key architectural contributions, and stores a structured Markdown summary directly into research-wiki/papers/. 2

A second test: run a bounded idea-discovery pass

Once the wiki and bridge are confirmed, test the core value proposition: generating a research hypothesis under cross-model scrutiny.

In Claude Code, provide a specific domain prompt:

/idea-discovery "mitigating catastrophic forgetting in low-rank adaptation fine-tuning"

The workflow executes three bounded steps:

Claude Code searches recent literature for established baselines and identifies unaddressed trade-offs.
The agent drafts a candidate hypothesis file under research-wiki/ideas/.
The skill calls Codex MCP to critique the proposal on empirical feasibility, missing control experiments, and prior art overlap.

Check the resulting idea card in research-wiki/ideas/. The document records both the proposed methodology and the reviewer's explicit counter-arguments. This dual perspective allows you to evaluate potential flaws before allocating GPU compute or writing code.

What hands-on users found

GitHub issue reports provide clear evidence of real-world friction and how maintainers have addressed user concerns.

In release v0.4.20, maintainers resolved issue #299, where the terminal spinner sequence wiped short single-line REPL responses. The terminal spinner had been clearing the entire active line upon completion, causing brief outputs such as confirmation messages to disappear. The fix ensures that Clear(UntilNewLine) only removes the trailing spinner characters while preserving rendered text. 3

In early September 2026, maintainers addressed issue #431, reported by contributor @oblivion-1521. Previous installer versions had automatically placed GitHub Copilot reviewer profiles into .github/agents/ for all projects, creating unnecessary configuration clutter for developers using Claude Code or Codex. The installer now only provisions these profiles when auto-review-loop is explicitly requested, and users can pass --no-agent-profiles to opt out entirely. 4

User inquiry #353 highlights configuration challenges with custom model allocations. A user with surplus Codex allocation asked how to configure Codex GPT as the main executor and GLM as the external reviewer, noting that default instructions assume Claude Code operates as the executor. The project supports ten alternative routing paths in docs/MODEL_COMBINATIONS.md, but navigating custom provider combinations requires manual editing of environment flags. 5

Issue #128, submitted by the creator of the Synapse research orchestration platform, evaluates ARIS's design principles. The feedback commends ARIS for driving complex research pipelines purely through Markdown instructions without heavy runtime daemons. At the same time, it emphasizes the importance of human-in-the-loop gates: automating background literature extraction and drafting is highly productive, but directional research decisions and GPU resource teardowns require human verification to avoid wasted compute cycles. 6

Use it now or keep it sandboxed

Adopt ARIS now when:

You use Claude Code and want an on-demand, local research assistant for paper ingestion and structured notes.
You have access to both Claude Code and Codex CLI, enabling cross-model critique to pressure-test research ideas.
You value a Markdown-first architecture that stores findings in transparent project repositories rather than closed databases.
You maintain personal oversight over experiment plans before triggering any remote execution.

Keep ARIS sandboxed when:

Your project involves proprietary codebase repositories or sensitive unreleased data.
You lack access to a secondary model endpoint, which disables the cross-model adversarial reviewer.
You expect fully autonomous, unsupervised GPU execution without human verification of compute allocation and cost boundaries.
Your local environment cannot run Python 3 or Node.js prerequisites.

The safest starting point is to install ARIS into an isolated scratch directory and use /alphaxiv and /research-wiki init to explore the documentation format before connecting remote compute resources.

Quick reference

Field	Details
Pick	`wanshuiyin/Auto-claude-code-research-in-sleep` (ARIS), a cross-model research skill pack. 1
Best for	Literature mapping, paper ingestion, hypothesis generation, and adversarial review loops. 1
Primary hosts	Claude Code (executor) and Codex CLI (reviewer via MCP bridge). 2
Install	Clone `~/aris_repo`, run `bash ~/aris_repo/tools/install_aris.sh`, and register the Codex MCP bridge. 2
Scope	82 Markdown-only skills across idea generation, experiment logging, and LaTeX paper writing. 1
License	MIT License for repository code and skill cards. 1
First test	Run `Ask GPT via codex MCP: what is 1+1?`, initialize `/research-wiki`, and test `/alphaxiv`. 2
Main caveat	Requires restarting Claude Code after MCP setup; custom model combinations require manual configuration. 2 5
Adoption signal	16.1k stars and active commits in September 2026. Weekly GitHub Trending did not list the repository. 1

References

1
ARIS repository README
github.com
2
ARIS Quick Setup Guide
github.com
3
Issue #299: Short REPL replies erased by spinner
github.com
4
Issue #431: Unwanted Copilot profiles deployed to projects
github.com
5
Issue #353: Model pairing and executor configuration
github.com
6
Issue #128: Research orchestration and human-steered execution
github.com