AI-Research-SKILLs MCP — 사용 사례, 설치 & 라이브 데모

왜 쓰나요

핵심 기능

87 skills across 22 categories: model architectures, tokenization, fine-tuning, post-training, distributed training, optimization, inference, safety
Autoresearch orchestrator skill manages full experiment lifecycle
References folder per skill with official docs, GitHub issues, tutorials
Covers modern post-training (TRL, GRPO, SimPO, verl, slime) and inference (vLLM, SGLang, TensorRT-LLM, llama.cpp)

라이브 데모

실제 사용 모습

ai-research-skill.replay ▶ 준비됨

0/0

설치

클라이언트 선택

~/Library/Application Support/Claude/claude_desktop_config.json · Windows: %APPDATA%\Claude\claude_desktop_config.json

{
  "mcpServers": {
    "ai-research-skill": {
      "command": "git",
      "args": [
        "clone",
        "https://github.com/Orchestra-Research/AI-Research-SKILLs",
        "~/.claude/skills/AI-Research-SKILLs"
      ],
      "_inferred": true
    }
  }
}

Claude Desktop → Settings → Developer → Edit Config 열기. 저장 후 앱 재시작.

~/.cursor/mcp.json · .cursor/mcp.json

{
  "mcpServers": {
    "ai-research-skill": {
      "command": "git",
      "args": [
        "clone",
        "https://github.com/Orchestra-Research/AI-Research-SKILLs",
        "~/.claude/skills/AI-Research-SKILLs"
      ],
      "_inferred": true
    }
  }
}

Cursor는 Claude Desktop과 동일한 mcpServers 스키마 사용. 프로젝트 설정이 전역보다 우선.

VS Code → Cline → MCP Servers → Edit

{
  "mcpServers": {
    "ai-research-skill": {
      "command": "git",
      "args": [
        "clone",
        "https://github.com/Orchestra-Research/AI-Research-SKILLs",
        "~/.claude/skills/AI-Research-SKILLs"
      ],
      "_inferred": true
    }
  }
}

Cline 사이드바의 MCP Servers 아이콘 클릭 후 "Edit Configuration" 선택.

~/.codeium/windsurf/mcp_config.json

{
  "mcpServers": {
    "ai-research-skill": {
      "command": "git",
      "args": [
        "clone",
        "https://github.com/Orchestra-Research/AI-Research-SKILLs",
        "~/.claude/skills/AI-Research-SKILLs"
      ],
      "_inferred": true
    }
  }
}

Claude Desktop과 같은 형식. Windsurf 재시작 후 적용.

~/.continue/config.json

{
  "mcpServers": [
    {
      "name": "ai-research-skill",
      "command": "git",
      "args": [
        "clone",
        "https://github.com/Orchestra-Research/AI-Research-SKILLs",
        "~/.claude/skills/AI-Research-SKILLs"
      ]
    }
  ]
}

Continue는 맵이 아닌 서버 오브젝트 배열 사용.

~/.config/zed/settings.json

{
  "context_servers": {
    "ai-research-skill": {
      "command": {
        "path": "git",
        "args": [
          "clone",
          "https://github.com/Orchestra-Research/AI-Research-SKILLs",
          "~/.claude/skills/AI-Research-SKILLs"
        ]
      }
    }
  }
}

context_servers에 추가. 저장 시 Zed가 핫 리로드.

claude mcp add ai-research-skill -- git clone https://github.com/Orchestra-Research/AI-Research-SKILLs ~/.claude/skills/AI-Research-SKILLs

한 줄 명령. claude mcp list로 확인, claude mcp remove로 제거.

사용 사례

실전 활용법: AI-Research-SKILLs

Fine-tune a Llama model with Unsloth + LoRA without Googling

👤 ML engineers running fine-tunes on a single GPU box ⏱ ~20 min intermediate

언제 쓸까: You have a dataset and want Claude to produce a runnable training script with correct flags.

사전 조건

Unsloth installed — pip install unsloth — the skill will remind you of the nightly build caveats

흐름

Describe the target run

Use the unsloth skill. Fine-tune Llama-3-8B on my dataset at ~/data/train.jsonl, QLoRA, 3 epochs, output to ./out.✓ 복사됨

→ Claude writes a script with correct Unsloth imports and real CLI flags
Ask for the config explanation

Walk me through each hyperparameter and why.✓ 복사됨

→ Reasoning grounded in Unsloth's docs, not generic ML clichés

결과: A training script that runs on the first try.

함정

Claude mixes HF Trainer and Unsloth idioms — Insist on 'use the unsloth skill only' — don't pull from Transformers skill

함께 쓰기: filesystem

Set up a vLLM server with correct tensor-parallelism for your GPUs

👤 Infra engineers serving LLMs in production ⏱ ~30 min advanced

언제 쓸까: You have 2-8 GPUs and want Claude to pick the right --tensor-parallel-size and --max-model-len.

흐름

State the hardware and model

Use the vllm skill. Serve Qwen2.5-72B on 4x H100s. Give me the exact launch command and sanity tests.✓ 복사됨

→ Correct TP size and quantization recommendation
Ask for load-test script

Now give me a locust or vllm-benchmark script to verify throughput.✓ 복사됨

→ Runnable benchmark using the right endpoint format

결과: A vLLM deployment with sanity checks and a benchmark baseline.

함정

Claude picks a TP size that doesn't divide the attention heads — The vLLM skill references list valid TP sizes per model family — have Claude cite that

함께 쓰기: aws

Draft an ML paper with correct structure and citation conventions

👤 Researchers writing for NeurIPS/ICML/ICLR ⏱ ~60 min advanced

언제 쓸까: You have experiment results and need a paper outline that matches the venue's style.

흐름

Provide results and venue

Use the ml-paper-writing skill. Target ICLR. Here are my results [paste]. Draft the intro, method, and experiments sections.✓ 복사됨

→ Structure follows venue conventions, ablation tables properly framed
Revise for reviewer concerns

What would Reviewer 2 push back on? Add preemptive responses.✓ 복사됨

→ Concrete weaknesses, not generic 'add more experiments'

결과: A draft that passes initial sanity review and is worth polishing.

함정

Claude over-claims in the abstract — Explicitly tell it to mirror the restraint of top-tier accepted papers

함께 쓰기: arxiv

조합

다른 MCP와 조합해 10배 효율

ai-research-skill + filesystem

Store training configs, logs, and checkpoints while Claude drives experiments

Save the run config under experiments/<date>-<name>/ and tail the logs.✓ 복사됨

ai-research-skill + arxiv

Pull related work while the ml-paper-writing skill drafts sections

Find 5 recent arXiv papers on GRPO and have the paper-writing skill weave them into the related work section.✓ 복사됨

도구

이 MCP가 노출하는 것

도구	입력	언제 호출
Autoresearch orchestrator	research goal	Complex multi-phase experiments
Fine-tuning (Axolotl, LLaMA-Factory, Unsloth, PEFT)	dataset + base model	SFT / LoRA / QLoRA runs
Post-training (TRL, GRPO, OpenRLHF, SimPO, verl, slime)	reward model or preference data	Alignment, preference optimization
Distributed training (DeepSpeed, FSDP, Megatron, Accelerate)	model + cluster topology	Multi-node or multi-GPU training
Inference (vLLM, SGLang, TensorRT-LLM, llama.cpp)	model + hardware	Serving a model efficiently
Optimization (Flash Attention, bitsandbytes, GPTQ, AWQ, GGUF)	model weights	Fitting a model on smaller hardware

비용 및 제한

운영 비용

API 쿼터: None for the skill itself
호출당 토큰: Heavy — SKILL.md + references can load 5-10k tokens per sub-skill
금액: Free — skills are local files; you pay for compute you run
팁: Scope prompts to one sub-skill at a time; the full 87-skill library is too big to load simultaneously.

보안

권한, 시크릿, 파급범위

자격 증명 저장: No credentials — skill is prompts + references

데이터 외부 송신: None from the skill itself

Some referenced frameworks have rapid API churn — confirm against the current version if a call fails.

문제 해결

자주 발생하는 오류와 해결

Claude blends two frameworks (e.g. Accelerate + DeepSpeed) incorrectly

Name one skill in the prompt and tell Claude to ignore the other.

References feel outdated

The repo is TeX-heavy and may lag the bleeding edge. For brand-new frameworks, supplement with a fresh fetch of the official docs.

Skill not auto-invoking on a relevant prompt

Mention the framework by name — 87 skills overlap and auto-routing is fuzzy.

대안

AI-Research-SKILLs 다른 것과 비교

대안	언제 쓰나	단점/장점
scientific-agent-skill	You need biology/chemistry/clinical, not ML training infra	Different domain focus
huggingface MCP	You want live HF Hub operations rather than expert prompts	MCP gives you real API actions; this skill teaches the patterns

AI-Research-SKILLs