AI Agents — the easy setup

Three agents, three jobs, one afternoon. This is the exact stack running the 19Keys operation: a coding agent in your terminal, a personal agent with skills, and an always-on agent that lives on a server and answers your messages. Set them up in order — each one builds on the last.

Requirements: a terminal you're comfortable opening (if not, start with the terminal-setup guide on this hub) and about an hour.

What an AI agent actually is

A chat model answers questions. An agent is a model wired to tools — it can read and write files, run commands, browse, and keep working toward a goal across many steps without you driving each one. The three below differ mainly in where they live: your terminal, your machine's background, or a server that never sleeps.

Tier 1 — Claude Code: the coding agent (start here)

The agent that builds things with you in the terminal. Made by Anthropic.

Install — macOS / Linux:

curl -fsSL https://claude.ai/install.sh | bash

Install — Windows (PowerShell):

irm https://claude.ai/install.ps1 | iex

First run:

cd your-project-folder
claude

It opens a session in that folder — ask it to build, fix, or explain anything there. Success looks like: a > prompt, and claude --version prints a version number.

Docs: claude.com/claude-code

Tier 2 — Hermes Agent: the open-source personal agent

Hermes Agent by Nous Research — an open-source agent CLI with bundled skills (research, files, coding, browsing) that you can read, modify, and own. The sovereign play: nothing about it is a black box.

Install (clone + venv — the dependable path):

git clone https://github.com/NousResearch/hermes-agent ~/.hermes/hermes-agent
cd ~/.hermes/hermes-agent
python3 -m venv venv
venv/bin/pip install -e .
ln -s "$PWD/venv/bin/hermes" ~/.local/bin/hermes   # put it on your PATH

(Check the repo README for the current one-line installer — the project moves fast.)

First run:

hermes

The first launch walks you through model-provider setup (bring your own API key — it lives in ~/.hermes/config.yaml, never in a repo). Success looks like: hermes --version prints a version, and the chat loop answers you.

Two files worth knowing in ~/.hermes/: SOUL.md (the agent's standing identity — edit it to make the agent yours) and config.yaml (models and keys).

Tier 3 — OpenClaw ("Clawd"): the always-on agent

OpenClaw is the agent that does not close when your laptop lid does. It runs as a background gateway on a machine that stays on (a Mac Mini is perfect), keeps persistent memory and identity files, and talks to you through the channels you already use — Telegram, WhatsApp, Discord. This is the "Jarvis that texts you back" tier.

Install (needs Node.js — brew install node if you don't have it):

npm install -g openclaw@latest
openclaw onboard --install-daemon

The onboarding wizard walks you through the gateway, workspace, channels, and skills, and installs it as a service (launchd/systemd) so it survives reboots.

Check it's alive:

openclaw gateway status
openclaw doctor

The pattern that makes it work: give the agent a home directory with identity files (AGENTS.md for working rules, an identity/soul file for voice, a memory file it appends to), give it ONE clear job to start, and put hard cost caps on the model spend before you let it run unattended.

Docs: docs.openclaw.ai · Getting started

Which one do you actually want?

The rules that keep agents useful

1. One job first. An agent with one clear job beats an agent with ten vague ones.
2. Identity files are the steering wheel. AGENTS.md / SOUL.md decide how it behaves when you're not watching. Write them deliberately.
3. Keys live in the environment, never in code or repos. Every project on this hub follows that rule; your agents should too.
4. Cost caps before autonomy. Set spend limits before you let anything run unattended overnight.
5. Always-on needs always-on hardware. A used Mac Mini on your shelf is the sovereign data center.

Part of KeyzHub — take the code, build your own.

Next rung: your agent should email you first — /agent-proactive-email.