Machines As Tools — Development skill for Claude Code
기기를 도구로 (Machines as Tools) — Claude Code 세션 하나가 남는 맥을 MCP 도구로 쓰는 구조.
How to install Machines As Tools
This entry records only its repository, not the path inside it, so there is no
exact command to give. Open joonlab/machines-as-tools and copy the folder into
~/.claude/skills/, or the file into ~/.claude/agents/.
What Machines As Tools does
기기를 도구로 (Machines as Tools) — Claude Code 세션 하나가 남는 맥을 MCP 도구로 쓰는 구조. 원격 실행기 mm + MCP 서버 multimac(도구 19개) + 실측 문서
Alternatives in Development
- MCP Builder — Guide for creating high-quality MCP servers to integrate external APIs and services 94.1k ★
- Mempalace Help — Show comprehensive MemPalace help — available skills, MCP tools, CLI commands, hooks, and architecture 58.6k ★
- Chrome Devtools MCP — Chrome DevTools for coding agents 29504 108 1 31.7k ★
README
machines-as-tools — 기기를 도구로
**English summary.** Three Macs cannot run one Claude Code session at the same time: a session is a single process appending to a single transcript. But Claude's reasoning runs on the server, and what actually eats a machine's RAM is the *tools* the session drives — browsers, builds, renders, scans. So this repo keeps **one session in charge** and registers spare Macs as **MCP tools** (19 of them, served by an MCP server called `multimac` on top of a shell wrapper `mm`). The rule of thumb is a cost ladder: deterministic shell work first (no LLM), a cheap lean worker with a short brief second, a stronger model third, and a full conversation fork only as a last resort. Measured on a laptop (M5 Max 128GB), an always-on home Mac (M3 Max 64GB) and a MacBook Air (M2 8GB), 2026-10-08 to 10-09; dollar figures are API-price equivalents.
Measurement Before → After Same summarization task: full conversation fork → ~1,000-char brief + Haiku worker $2.14 → $0.0146 (about 147×) Fan-out to 3 Macs: Workflow proxy agents → mm batch36 s, ~220k proxy tokens → 6.2 s, 0 proxy tokens Remote worker context: default config → lean config 95,209 → 26,934 tokens ($0.3119 → $0.0283) ssh round trip: new connection → ControlMaster reuse 0.2–0.3 s → 0.04 s
한 줄
**에이전트를 늘리기 전에, 기기를 도구로.** 세션은 하나로 두고, 남는 기기는 그 세션의 도구로 등록한다.
영상 (71초)
[](media/machines-as-tools_71s.mp4)
[media/machines-as-tools_71s.mp4](media/machines-as-tools_71s.mp4) — 71초짜리 한국어 컨셉 영상이다. 음악과 효과음이 깔리고 내레이션은 없다. 개인용 비공개 영상 킷(motion-reel)으로 만들었고, 리뷰어 두 명이 따로 검토한 뒤 고쳐서 품질 게이트를 통과했다. 이 영상도 이 저장소의 render 도구로 홈맥에서 렌더했다.
왜 만들었나
나는 노트북 한 대에 Claude Code 프로세스를 100개 넘게 띄워 두고 일한다. 그러다 무거운 작업을 한 기기에서 겹쳤다. 영상 렌더 워커 8개(약 9GB)와 병렬 `claude -p` 여러 개를 함께 돌렸고, 그 결과 커널 패닉을 두 번(2026-06-19, 2026-09-27) 겪었다. 그동안 집에 있는 64GB 맥과 맥북에어는 거의 놀고 있었다.
마침 공개 오픈소스 [OpenRig](https://github.com/mvschwar
Related Skills
Zengram
A Multi Agent Memory MCP That Connect Agents Across Systems and Machines
Dotenv Vault
Sync environment variables across machines for consistent Claude Code behavior.
Harness Remote
Native-session control plane for Codex, Claude Code, OpenCode, OMP and PI. Run, resume and hand off coding ses
Htb Writeups
The most comprehensive Hack The Box writeup collection - 500+ machines, 400+ challenges, interactive knowledge
Animation Pass
Review animation implementation, state machines, transitions, and gameplay readability.
Codex Claude Transfer
Local-first CLI (cct) that moves Codex & Claude Code sessions between machines: export to a .codexbundle, copy
Related Agents
Brain Merge
Semantically merge brain knowledge from multiple Claude Code machines. Use when brain-sync or brain-join detec
Journey Orchestrator
Use when the task requires designing cross-channel customer journeys, mapping touchpoint sequences, planning j
Business Logic Analyst
Business logic analysis specialist. Extracts business rules, explains business processes in natural language,