Karpathy method 6 AI CLIs Mac · Win · Linux

Your AI setup,
booted like an operating system.

AI-OS is the single source of truth for everything AI in your dev workflow. Spec → Verifier → Environment as real workflows, ~300 skills shared across every CLI, MCP servers as declarative YAML, all restored with one command.

$ npx @edd_remonts/create-ai-os what is this? →

// new Mac to full AI setup in ~5 minutes.

6
AI CLIs, one config
300+
skills, symlinked
10
MCP servers (YAML)
1
command to rebuild
// 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.

METHOD

Spec → Verifier → Environment

Karpathy's framework as 5 recurring workflows, 3 hard rule-sets, and 3 verifier gates applied after every task.

SKILLS

One source of truth, every CLI

14 required superpowers + ~300 curated & vendored skills, symlinked into Claude Code, Codex, Gemini, Antigravity, Hermes & MiniMax.

CONFIG

MCP + dotfiles as code

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

CI

Proven on 3 runners

macOS, Linux & Windows CI validate the install scripts on every PR. The setup is idempotent and has a dry-run mode.

// what it does

Three layers, symlinked top to bottom.

Skills and config flow down from a single repo into every CLI, and every CLI reads from the same shell & system underneath.

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 — Claude Code, Codex, Gemini, Antigravity, MiniMax, and Hermes, which reads the shared directory in place instead of getting a copy. Right, bottom: it generates the config for 10 MCP servers plus the local memory stack (FalkorDB, Ollama, and the AST and semantic code indexers) that those CLIs call into. Along the bottom: every non-trivial task runs Spec, Execute, Verify, Archive, then resets the spec.
The whole system on one page. SVG · PNG — generated with the vendored diagram-design skill, skinned from context/06_brand.md.
// layer 0: ai-os

The repo that owns everything

18 master instructions in CLAUDE.md, 7 context files, 6 workflows, 4 rules, 3 verifiers, flat global skills + optional ECC / claude.tools bundles, and a replicable dev-env. Change it here, it propagates everywhere.

// idempotent

Run it twice, safely

The installer never breaks on re-run. DRY_RUN=1 simulates the whole thing without touching your system.

// layer 1: clis

Six CLIs, one skill set

Claude Code · Codex · Gemini · Antigravity · MiniMax · Hermes. Each reads skills from its own home dir, but those dirs are symlinks back to AI-OS. One edit, six assistants updated.

// layer 2/3

Shell & system

zsh + Oh My Zsh + p10k + Warp on Mac, PowerShell + Windows Terminal on Windows, Homebrew / apt underneath.

// branded output

Diagrams that match the brand

27 editorial diagram types via the vendored diagram-design skill — self-contained HTML + inline SVG, dark-first, skinned from one token file. No Mermaid-slop, no shadows.

// system manual

Interactive System Internals & Manual

Click on any command in the sidebar to simulate execution and query detailed specs, configuration files, environment variables, and wiring methods of the AI-OS.

Terminal Menu
ai-os --status
➜ ~/ai-os
// command reference

Every command, and what it actually does.

Five surfaces: slash commands that work in any CLI, AI-OS skills you call by name, the ai-os shell CLI, the installers, and how each client enters a session.

Global slash commands

any CLI · ai-config/commands/ + matching skill
/action

Idea to reviewed code in one command: spec → plan → build → review, without stopping between phases.

/full

The complete delivery chain: spec → plan → build → review → code-simplify → ship.

/gen-html

Generate a rich, responsive, self-contained HTML document from a description — or convert an existing .md/.mdx file.

/saas-review

Whole-app design review: enumerate and visit every route, audit each screen against the expensive-UI criteria, report prioritised findings, then apply the approved fixes.

/sass-review

Alias of /saas-review — catches the common "SaaS" typo instead of failing.

/fix-ui-ux

Full Impeccable remediation pass: context + audit + critique, targeted repairs, then polish and document.

AI-OS skills, called by name

any CLI · ai-config/skills/
ai-os-quickstart

One-line bootstrap in a fresh CLI: verifies the 14 required superpowers are installed, runs setup, confirms the environment is ready.

ai-os-karpathy

Loads the full Spec + Verifier + Environment method plus persistent context — profile, projects, preferences, tools, sources.

ai-os-memory

Operate and inspect the memory stack: graph status, project indexing, semantic code search, graph browsing.

ai-os-loop

Build and run reusable loops — the 4-file system (TASK, INSTRUCTIONS, PROGRESS, outputs) with 6 blast-radius safety levels.

diagram-design

27 editorial diagram types as self-contained HTML + inline SVG, skinned from context/06_brand.md. The diagram above came from it.

The ai-os CLI

shell · ~/.local/bin/ai-os
ai-os bootstrap [--refresh]

First-time or on-demand full-stack setup, then auto-index every project.

ai-os memory status

FalkorDB + Ollama + MCP health with node and edge counts.

ai-os memory start | stop | logs

FalkorDB container lifecycle. stop preserves the data; logs tails it.

ai-os memory query "<q>" [path]

Semantic search across indexed code via grepai.

ai-os memory search "<q>" [path]

Same, but falls back to the AST index when grepai isn't running.

ai-os memory reindex [path]

Refresh the codebase-memory-mcp AST index.

ai-os memory browse | visualize

Open the FalkorDB web UI, or every memory UI plus grepai status at once.

ai-os loop ls | add <dir> | rm <id>

List loops, register one (needs TASK.md + LOOP_INSTRUCTIONS.md), or mark one removed while keeping its directory.

ai-os loop status | show | run | pause <id>

Inspect a loop's files, level and last activity; cat its PROGRESS.md; open a terminal at it; or disable its scheduling.

Install & verify

shell · from the repo root
npx @edd_remonts/create-ai-os

Bootstrap from nothing — clones the repo and runs the right installer for your OS.

bash setup/install-mac.sh

The full Mac install. Idempotent — safe to re-run. Honours SKIP_BREW, SKIP_NPM, SKIP_PIP, SKIP_MCP and friends.

powershell -File setup/install-windows.ps1

The Windows equivalent, same guarantees.

bash setup/install-mac.dry-run.sh

Simulates the entire install against a temp HOME, touching nothing real. This is the CI gate on every PR.

bash setup/verify.sh

Verifies a live install: 19 required checks (skills, adapters, superpowers, per-machine env isolation) plus optional ones that warn instead of failing.

bash setup/install-ecc.sh

Wires the vendored ECC bundle — 271 skills, 28 hooks, agents — into the core CLIs.

bash setup/install-claude-tools.sh [--check]

claude.tools + gstack skills and the Codex plugin. --check validates structure only, for CI.

Entering a session

shell · per client
claude  ·  codex  ·  gemini

Nothing to pass. Each client's global instruction file is a symlink to the rendered AI-OS bridge, so context loads on start.

hermes chat --skills ai-os-quickstart

Hermes takes skills as a flag; it reads the shared skills directory natively via skills.external_dirs.

cat CLAUDE.md

The escape hatch for any client without a global instruction file — paste it at the start of the conversation.

// the method

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

6 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 (12 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: +271 ECC skills
  • setup/install-claude-tools.sh: +12 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.

45 CLI tools & runtimes

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

22 Apps & fonts

VS Code, Warp, iTerm2, Windsurf, Docker Desktop, Chrome, Firefox, Raycast, TablePlus, Postman, KeePassXC + 4 Nerd Fonts… Homebrew casks

43 Global Node packages

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

28 Python packages

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

~180 AI skills

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

+ Services & shell

Memory stack (FalkorDB + Ollama + Graphiti + code indexers via Docker), 5 global commands (/action /full /gen-html /saas-review /fix-ui-ux), MCP config, Oh My Zsh + Powerlevel10k, dotfile symlinks. background layers

// end of boot

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