Overview
AREX-Skill is a library of executable Agent Skills distilled from public machine-learning repositories, together with DisCo, the CLI agent that does the distilling. The premise is that most research knowledge is written for humans: papers explain methods, repositories hold implementations, and blogs record the tricks and failure modes, leaving an agent to search, decide what applies, assemble a workflow, debug the missing steps and then judge whether the result is credible. AREX-Skill turns that operating knowledge into something an agent can load directly.
A skill here is not a repository summary. It is a self-contained unit that states when a capability applies, what to run, how to validate the result, and how to recover when an experiment fails. Skills use the open Agent Skills format as their packaging convention — a SKILL.md with optional references/ and scripts/ directories — so they are portable across harnesses; the project documents integration with Codex, Claude Code and Pi among others.
When a repository exposes several capabilities, its skills are composed into a repository skill graph, and a router narrows a request to an area, family, repository and workflow before the agent loads only the branch it needs. That progressive disclosure is what keeps the initial context small while leaving deeper instructions reachable. As of the 2026-08-27 library update the published collection covers 20 research areas, 178 package families, 1,000 repositories and more than 5,000 verified skills.
What it does
- 5,000+ verified, executable skills distilled from 1,000+ machine-learning repositories across 20 research areas and 178 package families
- Skills packaged in the open Agent Skills format (SKILL.md plus optional references/ and scripts/), portable to Codex, Claude Code, Pi and other harnesses
- Every skill records scope, routing, the workflow to run, validation checks, and recovery guidance for failed experiments
- Repository skill graphs with a router that narrows a request to one branch, so the agent loads only the instructions the task needs
- The DisCo CLI (published on npm as @arex-skill/disco) distils a repository into skills, with Creator and Researcher workflows
- Skills are maintained against upstream changes, with source commits, validation steps and refresh requirements documented per repository
Getting started
There are two ways in: browse and install skills from the published library, or run DisCo to distil a repository of your own. The repository's library guide and architecture guide cover the runtime and routing model.
Get the library
Clone the repository; the skills live under skills/ with a library guide explaining the routing layout, and docs/repository-catalog.md lists the full area and package-family inventory.
git clone https://github.com/VectorSpaceLab/AREX-SkillInstall the DisCo CLI
DisCo is the agent that turns a repository into verified, executable skills. It is published on npm.
npm install -g @arex-skill/discoLet the router pick the branch
A research agent starts from a concrete goal, routes the request to the relevant skill graph, loads only the needed SKILL.md branch, runs the recommended procedures and scripts, and validates the result against the skill's checks — following the recovery guidance and iterating when a run fails.
Commands and code are distilled from the project's own documentation — always check the official repo for the latest.
When to use it
- Give an autonomous ML research agent operating knowledge of a library it has never used, including how to validate the run
- Turn your own repository into skills your coding agent can route to, rather than pasting the README into every prompt
- Keep context small on a long research task by loading one branch of a repository skill graph instead of a whole documentation set
- Standardise how a team packages agent instructions, using the open Agent Skills format rather than per-harness conventions
How AREX-Skill compares
AREX-Skill alongside other open-source agent skills & plugins tools AI/TLDR tracks, ranked by GitHub stars.
| Tool | Stars | What it does |
|---|---|---|
| Superpowers | ★ 304k | A composable skills plugin that installs a spec-first, TDD, subagent-driven development methodology into Claude Code, Codex, Cursor, Gemini CLI and other coding harnesses. |
| Skills for Real Engineers | ★ 301k | Matt Pocock's everyday agent skills for coding agents, covering alignment grilling, planning, code review and research — small, composable and meant to be edited. |
| ECC | ★ 287k | An installable plugin that adds 68 agents, 286 skills, hooks, rules, memory and an agent-config security scanner to Claude Code, Codex and other coding harnesses. |
| Karpathy Coding Guidelines | ★ 223k | A Claude Code plugin and CLAUDE.md rule set built from Andrej Karpathy's observations on LLM coding pitfalls: think before coding, keep it simple, make surgical changes, and work to verifiable success criteria. |
| Ponytail | ★ 174k | A skill and plugin that makes a coding agent take the laziest correct route — a seven-rung ladder from "does this need to exist?" to "the minimum that works" — while keeping validation, security and accessibility off the chopping block. |
| Caveman | ★ 114k | A skill plus local proxy that shortens an AI coding agent's prose while leaving code, commands and paths untouched — the rule file cuts what the agent writes, the proxy compresses what it reads. |
| Addy’s Agent Skills | ★ 110k | A pack of production engineering skills for AI coding agents, with nine lifecycle slash commands — /spec, /plan, /build, /test, /review, /ship — installable into 70+ agents. |
| AREX-Skill | ★ 341 | Repositories distilled into executable skills an ML research agent can route to |