Karpathy method 6 AI CLIs Mac · Win · Linux

Your AI setup,
booted like an operating system.

One command restores your context, 186 skills and 10 MCP servers across six AI CLIs, on a brand-new Mac.

$ npx @edd_remonts/create-ai-os see it ship a real product →
6
AI CLIs, one config
186
skills, symlinked
10
MCP servers (YAML)
3
OS runners green in CI
// what is it

The answer to "I just got a new Mac. How do I get my AI setup back?"

Most devs who use AI assistants daily end up with skills scattered across ~/.claude/, ~/.codex/, MCP servers hardcoded with no version control, and the Spec→Plan→Execute→Verify loop living only in their head. AI-OS puts all of it in one git repo, and keeps the products it builds out of it once they outgrow it.

METHOD

Spec → Verifier → Environment

Karpathy's framework as recurring workflows, hard rule-sets, and verifier gates applied after every task. One active Spec at a time; nothing ships without a runtime check.

SKILLS

One source of truth, every CLI

14 required superpowers + 172 curated & vendored skills, symlinked into Claude Code, Codex, Gemini, Antigravity, Hermes & MiniMax. Plus 16 more hand-picked from a 272-skill ECC snapshot.

CONFIG

MCP + dotfiles as code

10 MCP servers as declarative YAML, generated at install. zsh, Warp, Brewfile & git config versioned in the same repo.

BOUNDARY

Framework above, products below

Everything above apps/ is the operating system. What it builds lives under apps/ until it needs its own release cadence, then it splits out.

Whole-system architecture. Left: the ai-os repo holds context, rules, workflows, verifiers, ai-config and dev-env. Centre: setup/install-mac.sh reads it as the single entry point. Right, top: it symlinks skills and renders the instruction adapter into six CLIs. Right, bottom: it generates the config for 10 MCP servers plus the local memory stack. Along the bottom: every non-trivial task runs Spec, Execute, Verify, Archive, then resets the spec.
The whole system on one page. SVG — generated with the vendored diagram-design skill, skinned from context/06_brand.md.
Plansmith, a live product built with AI-OS: a search and filter UI over a corpus of 900+ real product-idea plans, each scored for money, learning and fun potential.

The method's proof isn't a demo. It's a product that outgrew the framework.

Plansmith started as one more directory under this repo's apps/: a scraper, a corpus, a formatter, a daily pipeline, a Vue explorer. It ran the same Spec → Verifier → Environment loop as everything else here, one 30-minute block at a time, until it was large enough to need its own release cadence.

beforeOne repo, 2 deployed products, corpus growing daily under apps/data/.
splitA same-day extraction: fresh history, same internal paths, verified green before a single file moved.
nowA separate repo, its own deploy pipeline, 3 live services — and AI-OS is one product lighter.
visit plansmith →

Every non-trivial task runs the same loop.

Spec quality first, verifier passes second, runtime evidence third. No "it builds, ship it."

Spec

Define the objective

One active spec: objective, acceptance criteria, non-goals, risks.

Plan

Break into blocks

Blocks of ≤30 min, dispatched in parallel when independent.

Execute

Block by block

Verify after each block. Sub-agents self-report; you check the handles.

Verify

Gates per task

Quality checklist, self-critique, source-check. Evidence before "done".

// install

One command. Verify.

Pick your OS. Copy the block. Most of the 30–60 min first install is just downloading packages.

# 1. bootstrap (clones + runs the Mac installer)
npx @edd_remonts/create-ai-os
cd ai-os

# 2. verify (20 health-check sections)
bash setup/verify.sh

Prerequisites

  • Apple Silicon / Intel Mac, Windows 10-11, or Linux
  • git, curl and an internet connection
  • ~30–60 min for the first run (mostly downloads)

Optional layers

  • setup/install-ecc.sh: the full 272-skill ECC snapshot as a Claude Code plugin
  • setup/install-claude-tools.sh: 12 more claude.tools / gstack skills
  • DRY_RUN=1: simulate without touching your machine
// full manifest

Everything it puts on your machine.

The installer prints this list and asks for confirmation before touching anything. Nothing is installed silently.

70 CLI tools & runtimes

git, gh, node, python, uv, go, rust, docker, kubernetes-cli, terraform, ripgrep, fzf, bat, eza, lazygit, ollama, helm… Homebrew

43 Global Node packages

claude-code, codex, gemini-cli, MCP servers, typescript, vite, eslint, prettier, biome, turbo, nx… via pnpm

30 Python packages

ruff, black, mypy, pytest, ipython, pandas, numpy, httpx, typer, bandit, pip-audit, poetry, mcp… isolated uv venv

186 AI skills

Symlinked into every CLI (Claude, Codex, Gemini, Hermes, Antigravity, MiniMax) plus the 14 required superpowers. ai-config/skills

9 Global commands

/action /full /gen-html /saas-review /fix-ui-ux and 4 more, in every CLI that supports slash commands. ai-config/commands

+ Services & shell

Memory stack (FalkorDB + Ollama + Graphiti + code indexers via Docker), MCP config, Oh My Zsh + Powerlevel10k, dotfile symlinks. background layers

Reboot your workflow.

Stop reconfiguring six CLIs by hand. Put the method, the skills and the environment in one repo, and bring it back in one command.

◆ github.com/eddremonts86/ai-os ↑ back to top