Kostya — this is the upgrade.

The chat window is a demo. The CLI is the workshop.

Last night I did a garbage job explaining this, so here's the clean version. In the desktop app you describe an app and Claude builds you something gorgeous — you've been doing exactly that, and it shows. But every one of those apps lives in a sandbox you have to copy things out of. You're the courier between the AI and your real files, your real projects, your real business.

The command-line version deletes the courier. It runs in your actual folders. It edits real files, runs real commands, spins up background agents, and connects the apps you already built to each other through code. Same Claude brain — but now it has hands, and the hands are in your workshop instead of behind glass.

“Is it safe to let this thing into my computer?”

Fair question — and the honest answer is: only if you set it up so it can't surprise you. Here's exactly how the trust works, so you're never handing over blind control.

Bottom line: the CLI is more auditable than the app, not less — you can see every command it wants to run and veto it. The skepticism is correct; it just points at the installer, not the tool.

Install the CLIs

You liked the idea of flipping to another Warp tab and working in Codex — so here's both. Run Claude Code in one tab, Codex in another, and you've got two AI coworkers side by side. Every command below is copied verbatim from the official docs (verified 2026-07-23).

Claude Code (Anthropic) — my daily driver

# official native installer — macOS / Linux
curl -fsSL https://claude.ai/install.sh | bash

Prefer Homebrew? brew install --cask claude-code. Prefer npm (needs Node 22+)? npm install -g @anthropic-ai/claude-code. Then start it by running claude inside any project folder. Needs a Claude Pro/Max account to sign in.

Codex (OpenAI) — the other tab

# official — via npm
npm install -g @openai/codex

# or Homebrew
brew install --cask codex

# or standalone installer
curl -fsSL https://chatgpt.com/codex/install.sh | sh

Start it with codex, then choose Sign in with ChatGPT. Same move as Claude — launch it inside the folder you want it to work in.

# import Anthropic's public signing key + show its fingerprint
curl -fsSL https://downloads.claude.ai/keys/claude-code.asc | gpg --import
gpg --fingerprint security@anthropic.com

# the fingerprint MUST read exactly:
# 31DD DE24 DDFA B679 F42D  7BD2 BAA9 29FF 1A7E CACE

Your first five minutes

This is the whole on-ramp — open Warp, and step through it. Nothing here changes a file without your yes.

Notice the shape: it explains, it asks, then it acts. That “Allow?” prompt is the safety model from earlier, live.

Four things that make it a machine

These are the four capabilities that took me from “nice code helper” to “this runs part of my business.” Open each one. The prompt at the bottom of this page will go research all of them against your own setup — but skim them first so you know what you're aiming at.

# try this inside Claude Code:
Create a skill called /gig-sheet that turns a booking
email into a one-page run sheet: date, venue, load-in
time, audience, tech notes, and my set list.

# try this — point it at something you built:
Take one app I've already built. Show me the smallest
change that uses the Claude Agent SDK to make it act on
its own data instead of just displaying it.

# try this:
What MCP servers exist for the tools I use — calendar,
email, database, Slack? Show me how to connect one, and
explain exactly what it can and can't touch.

# try this:
Give me one realistic morning scheduled-agent for my
business — does something useful with my real data and
leaves the result for me to approve, sends nothing itself.

Desktop app vs. the CLI

I'm not going to pretend the app is bad — it's where you'll still prototype UIs. But as a daily driver, this is the honest comparison:

Now let it study your setup

Here's the payoff. Open a fresh Claude Code session inside your main projects folder and paste this. It researches everything above, then looks at your actual machine — safely, read-only, asking before anything — and hands you a personal report of where the CLI and SDK would save you the most time. It ends by offering to build one small starter so you learn by running something real.

You are my onboarding guide to Claude Code and the Claude Agent
SDK. I'm brand new to the CLI but I build apps in the Claude
desktop app and I run a live-magic / mentalism business. Work in
phases and DO NOT change, create, or delete any file, and DO NOT
run any command that modifies anything, without stopping to
explain it and getting my explicit yes first. Start read-only.

PHASE 1 — RESEARCH (no file access needed): In plain language,
teach me the four capabilities that make the CLI more than a chat
window: (1) Skills & slash commands, (2) the Claude Agent SDK,
(3) MCP servers, (4) hooks / scheduled agents / subagents. For
each: what it is, one concrete example for a magic-performance
business, and where the official docs are.

PHASE 2 — DISCOVER MY SETUP (read-only): With my permission for
each step, look through this folder and my projects to figure out
what I actually have — languages, frameworks, the apps I've built,
databases, calendars, services. Don't open anything sensitive
without asking. Summarize my stack back so I can confirm it.

PHASE 3 — ANALYZE: Give me a prioritized list of where moving to
the CLI + Agent SDK would help me most. For each: what it is,
which capability it uses, rough effort, and the concrete payoff
for my business. Be honest about where the desktop app still wins.

PHASE 4 — BUILD A STARTER (only after I pick one): Let me choose
one item. Scaffold the smallest working version — a simple skill
or a small MCP/SDK hookup — explaining each step, asking before
you create or run anything, so I understand what we built.

Throughout: explain like I'm sharp but new to this, keep the
safety rails visible, never take an action I haven't approved.

The copy button grabs the full prompt (it's a bit longer than what's shown). Paste, hit enter, and go make coffee.

Run your own private AI at home

Remember the “Spark boxes” I mentioned setting up at the office? That's the other half of this. A Spark box is just a dedicated, always-on machine running an AI model locally — your own private ChatGPT that never phones home. I reach mine from anywhere over a private mesh. You can build the exact same thing, and for a mentalist it's genuinely valuable: your client data, your booking info, and — the big one — your methods and routine ideas never leave your house or land in some vendor's logs.

The one thing not to buy

Your hardware options, honestly

Do it this weekend — free, on the Mac you have

This is the whole pattern — a private model with a real API — in about fifteen minutes and zero dollars. Each command is safe and reversible.

# 1. install Ollama (the easiest local-AI runner)
brew install ollama

# 2. pull a great small model. On 32GB+ RAM, try qwen3:32b instead
ollama pull qwen3:8b

# 3. chat with it — fully private, offline-capable
ollama run qwen3:8b

# 4. prove it's a real API your own apps can call
curl http://localhost:11434/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{"model":"qwen3:8b","messages":[{"role":"user","content":"Say hi"}]}'

Want to reach it from your laptop at a gig? Install Tailscale on the home Mac and your laptop, sign into both with the same account, and they join one private network — no router settings, nothing exposed to the internet. That's the “from anywhere” part of my setup, and it's another fifteen minutes.

Where it earns its keep for you

• Private client email & booking work — draft quotes and follow-ups with real names and fees, none of it touching a cloud.
• A “second brain” over your own material — point it at your notes, scripts, and past shows: “what did I perform for the Hendersons in 2024?”
• Brainstorm methods & routines with zero disclosure risk — the reason this matters most for a mentalist.
• Prep offline — tighten a set, write emcee intros, in a green room with no signal.

I want to run a private local LLM on this Mac — my own
always-on AI that keeps my data at home. I'm new to this and
cautious about what gets installed, so explain each step and
ask before installing or running anything. Walk me through:
checking this Mac's memory and which model fits, installing
Ollama and pulling the right model, confirming it works + how
to call its local API from my apps, and optionally Tailscale
so I can reach it from my laptop at a gig without exposing
anything. Then suggest a few ways a mentalist could use it.