# AREX-Skill

> A library of 5,000+ executable agent skills distilled from 1,000 ML repositories, plus the DisCo CLI for routing and creating them.

AREX-Skill is an open-source project from VectorSpaceLab that turns knowledge from GitHub repositories and papers into executable, agent-readable skills for autonomous machine learning research. It ships a library of 5,000+ skills distilled from 1,000 ML repositories, along with DisCo, a CLI that routes tasks to relevant skills and can build new ones. The latest release listed is v0.2.1.

## What It Is

An AREX Skill packages operating knowledge for a coding agent: 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 and are organized around a SKILL.md file with optional references and scripts folders. When a repository exposes several capabilities, its skills are composed into a repository skill graph.

## How the Workflow Runs

A router narrows a request to an area, family, repository, and workflow, and the agent loads only the needed branch of the skill graph. The agent then runs the recommended procedures and scripts, validates results against the skill's checks, and follows recovery guidance if a run fails. DisCo's default Researcher mode loads and routes the library natively, and selected repository skills can be exported with a scoped router to other coding agents such as Codex and Claude Code.

## Building and Refreshing Skills

The library is produced by DisCo Creator through a four-stage distillation workflow: scope capabilities, ground them in evidence, construct a skill graph, and verify and refine it. Creator workflows can also build skills from papers and refresh existing repository skills against upstream changes. The library spans 20 research areas and 178 package families, including LLM serving, computer vision, biomedical AI, and robotics.

## Reported Benchmark Results

The project reports that Codex with AREX-Skill improved over Codex alone on MLE-bench, PaperBench, FrontierCS, and PassNet, with the agent setup and budget held fixed. For example, it lists the MLE-bench Any Medal rate rising from 31.11% to 72.89%. A full technical report is noted as forthcoming.

## Setup Path

DisCo requires Node.js 22.19.0 or newer and installs through a curl installer or npm. After configuring a model provider, run `disco repo-skills install` and then `disco`. Each skill carries its own license in its SKILL.md, which is authoritative for that skill; repository-level materials are Apache-2.0.

## Features
- 5,000+ executable skills distilled from 1,000 ML repositories
- Repository skill graphs with a router for progressive disclosure
- DisCo CLI with Researcher and Creator modes
- Export of selected skills to Codex and Claude Code
- Skill creation from repositories and papers with verification
- Refresh workflow for updating skills against upstream changes
- SKILL.md format with references and scripts

## Integrations
Codex, Claude Code, Pi, OpenAI, Anthropic, Gemini, Agent Skills format

## Platforms
WINDOWS, MACOS, LINUX, API, CLI

## Pricing
Open Source

## Version
v0.2.1

## Links
- Website: https://github.com/VectorSpaceLab/AREX-Skill
- Documentation: https://github.com/VectorSpaceLab/AREX-Skill/blob/main/docs/README.md
- Repository: https://github.com/VectorSpaceLab/AREX-Skill
- EveryDev.ai: https://www.everydev.ai/tools/arex-skill