Your personal AI stack, from zero: what to install and in what order

TL;DR

A personal AI stack has five parts: a host that runs on your machine, a model that does the reasoning, plugins that add capabilities, skills that tell the agent how to use those capabilities, and memory that persists between sessions. Install them in that order. Most people who get stuck setting up a personal AI are not stuck on a command, they are stuck because nobody told them the five parts existed.

If you have tried to set up a personal AI and given up, the problem was probably not technical. The instructions you found assumed you already knew what the pieces were. You installed something, it asked for an API key, you were not sure which kind, and there were four other things in the readme you had no model for.

So start with the model, not the commands.

What a personal AI stack actually is

Five parts. Each one does a job the others cannot.

The distinction people find most useful is between plugins and skills. A plugin is capability: it is the wiring that lets the agent call something. A skill is judgment about that capability: when to reach for it, what to ask, what a good result looks like. A plugin without a skill is a tool your agent owns and never picks up.

The other thing worth saying plainly: none of this requires you to know what a tool call is. You will talk to the agent in ordinary sentences. The layers exist so that you do not have to think about them.

Step 1: install the host

The host is the thing that actually runs. Everything else plugs into it, and your memory and plugin state live inside it, which is why this is the one decision that is annoying to reverse later. Pick it deliberately and then stop shopping.

This guide uses OpenClaw, an open-source assistant that runs locally on your own machine rather than as a hosted service.

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

It runs on macOS 15 and later (universal binary), Windows 10 20H2 or later and Windows 11, and Linux.

One thing that removes a common blocker: the installer brings Node itself, so you do not need to install Node first. If you have been putting this off because a previous guide opened with a runtime prerequisite, that step is gone.

Then onboard:

openclaw onboard

Onboarding is where you connect the surfaces you want the agent to work in. OpenClaw operates across WhatsApp, Telegram, Discord, Slack, iMessage and Signal, and can work with your email, calendar and files, and run commands on your machine. You do not have to connect all of them now. Connect one, ideally the messaging app you already have open all day, because an assistant you have to remember to visit is one you will stop visiting.

Step 2: give it a brain, and decide what it costs

At this point you have a host that cannot think. It needs access to a model, and this is the step most setup guides skip past, which is unfortunate because it is the step with the actual commitment in it.

You have two paths.

Bring a key for a hosted model. OpenClaw works with Claude, GPT, or a compatible model, and you paste in a key from whichever provider you choose. This is the higher-quality path and it is the one that costs money, billed by your provider according to how much you use.

Or run a model locally. OpenClaw supports local models, MiniMax 2.5 among them. This path costs nothing to run and asks more of your hardware.

Do this before you install anything else. A host with no model access will fail in ways that look like installation problems and are not, and you will spend an evening debugging the wrong layer.

Step 3: install exactly one plugin

Here is the part where people lose the thread: they finish onboarding, feel the momentum, and install six plugins in a row. Then nothing works and there is no way to tell which layer broke.

Install one. Choose one that produces a visible result on the first try, because your second install is decided entirely by whether the first one did something you could see.

A reasonable first pick is a shopping plugin, since the result is concrete and you can judge it immediately. SIL is 4GPTs' commerce plugin for OpenClaw, built on UCP (the Universal Commerce Protocol), and it is open source under Apache-2.0.

# ClawHub (recommended)
openclaw plugins install clawhub:@4gpts/sil

# npm
openclaw plugins install sil-openclaw

Then the first run is two plain sentences, spoken to your agent, not typed into a terminal:

1. Tell your agent:  "register me on sil"          (one browser sign-in, done)
2. Tell your agent:  "find me a mechanical keyboard under $100"

That second line is the best illustration of what a plugin actually is. You did not call an API or learn a command. You stated an intent, the agent loaded the plugin's bundled skill because the intent was a commerce one, and it went and searched, compared, pulled up product detail, and handed back links you can buy from.

If you want to see what that reasoning looks like in detail before you install anything, what an AI agent reasons about when it buys ski boots walks through a full run.

What skills and memory do, and why they come last

You do not configure these on day one. You notice them on day three.

Skills arrived with your plugin. You did not write them, and the reason "find me a mechanical keyboard under $100" worked without any further instruction is that the plugin brought its own guidance about how to handle a request like that. When you evaluate a plugin later, this is the quality difference worth looking for: not how many functions it exposes, but whether it knows when to use them.

Memory is what stops you repeating yourself. The value shows up the third time you ask for something in a category, when the agent already knows what you rejected last time and why. It is also the reason your host choice matters more than your plugin choices, since this is where that history accumulates.

Both of these are day-three concerns. Guides that open with them are the reason setup feels abstract.

When it does not work

Three failures cover most of it, and each one looks like something else.

The agent responds but never uses the plugin. Most likely you never stated an intent that triggers it. Plugins load on intent, not on launch. Ask for the thing the plugin does, in the words you would use with a person.

Everything fails right after install. Check the model layer before the plugin layer. A missing or invalid model key surfaces as generic breakage, and it is the most common cause of an evening lost to debugging the wrong thing.

The install finished but nothing responds. You probably have not run openclaw onboard. Installing the host and connecting it to a surface are two separate steps, and the first one succeeding tells you nothing about the second.

The general rule: work down the layers in the order you installed them. Host, then model, then plugin. Almost every confusing failure is a lower layer than the one where you noticed it.

The point of all this

The useful agent is the one you stop having to think about. Everything above is scaffolding for that: you set up five layers once so that afterwards you can say what you want in a sentence and stop managing the machinery.

Which is also the answer to "how much of this do I need to hold in my head". You do not. Tell your agent.

FAQ