I Tried Installing OpenClaw on Windows in an Hour. Here's What Actually Happened.

This is a submission for the OpenClaw Writing Challenge

This is a submission for the OpenClaw Challenge — Wealth of Knowledge prompt.

I saw the OpenClaw Challenge on DEV yesterday and decided to give it an hour. I'm a CS student on Windows, no prior Claude Code experience, currently in placement season and desperate for anything that triages my inbox. Spoiler: I did not get a working Gmail triage in an hour. But I learned enough about how OpenClaw works that I think it's worth writing down — especially for other Windows-first students who'll hit the same walls.

This isn't a success story. It's the honest first-hour walkthrough I wish I'd found before starting.

What OpenClaw actually is

OpenClaw is a personal agent framework. You install it, give it "skills" (Markdown files that describe what it should do and when), and it uses an LLM to reason through tasks. The pitch that pulled me in: a skill is just a SKILL.md file. No code, no deploy, no API gymnastics. You write Markdown, the model handles the rest.

That pitch is true. But getting to the point where you can actually write and run a skill takes more setup than the DEV announcement suggests — at least on Windows.

The install

I opened PowerShell and ran:

iwr -useb https://openclaw.ai/install.ps1 | iex
~~~{% endraw %}

It worked. I got a welcome banner and a nice lobster ASCII logo. The installer did catch that I was on Windows and showed this:

> Windows detected — OpenClaw runs great on WSL2!
> Native Windows might be trickier.
> Quick setup: {% raw %}`wsl --install`{% endraw %} (one command, one reboot)

I missed this warning. Pushed past it. In hindsight: read the warning. If you're reading this post and you're on Windows, **install WSL first.** The rest of this story is partly a case study in what happens when you don't.

## The dependency I didn't know about

I got to the auth step and hit:{% raw %}

Error: Claude CLI is not authenticated on this host.
Run claude auth login first, then re-run this setup.

This surprised me. I hadn't realized **OpenClaw uses Claude Code as its LLM backend**. Claude Code is Anthropic's own terminal CLI, and it's a separate install with its own auth flow. So before I could use OpenClaw, I had to:

1. Install Claude Code: {% raw %}`irm https://claude.ai/install.ps1 | iex`
2. Run `claude` and log in via browser (requires a paid Claude plan — Pro, Max, Team, or Console credits)

I already had Claude Pro from doing coursework, so this wasn't a blocker. But if you don't have a paid Claude account, **you can't use OpenClaw this way.** That's a real gate nobody mentioned in the challenge announcement.

Worth knowing up front.

## Setup finished — then broke again

After Claude Code was logged in, I re-ran `openclaw setup`. It finished cleanly:

Wrote ~/.openclaw/openclaw.json
Workspace OK: ~/.openclaw/workspace
Sessions OK: ~/.openclaw/agents/main/sessions

Great. I ran {% raw %}`openclaw tui` to start chatting and got:

Error: Missing gateway auth token.
Fix: set OPENCLAW_GATEWAY_TOKEN/OPENCLAW_GATEWAY_PASSWORD, pass --token/--password,
or resolve the configured secret provider for this credential.

OpenClaw has a gateway architecture — there's a long-running process (the gateway) that the CLI connects to. You have to start the gateway separately, generate a token, then pass that token to the client. Straightforward in concept, but the setup flow doesn't walk you through it.

This was the wall I didn't get past in my hour.

## What I did learn by poking around

The one command that was genuinely illuminating: {% raw %}`openclaw skills`.

It lists every skill OpenClaw knows about. On a fresh install, I counted **52 bundled skills**, 8 of them ready to use out of the box. Highlights:

- `taskflow-inbox-triage` — an email triage example already built by the OpenClaw team
- `gog` — Google Workspace (Gmail, Calendar, Drive) integration
- `himalaya` — IMAP/SMTP email via CLI
- `skill-creator` — a skill that helps you write other skills
- `github`, `notion`, `obsidian`, `slack`, `discord` — standard productivity integrations
- `camsnap`, `openhue`, `sonoscli` — IoT stuff I didn't expect

Most are marked "needs setup" — meaning the skill file is there, but you need to install the underlying CLI tool and authenticate it. It's a plugin model: OpenClaw orchestrates, the real work happens in well-established CLIs like `gh`, `himalaya`, `obsidian-cli`.

That's actually a smart design. Instead of reimplementing integrations, OpenClaw wraps tools that already exist.

## Three things I genuinely learned

**1. A "skill" is a thin layer over an existing CLI.**
When I thought "skill," I pictured a self-contained script. It's not. A skill is mostly a `SKILL.md` that tells the model *how and when* to call an underlying command-line tool. Want Gmail triage? You don't write mail parsing — you configure `gog` or `himalaya`, and your skill just tells the model which commands to run against it.

This means the Markdown-only pitch is half true. The Markdown is your job. But the CLI underneath still needs to exist and be authenticated.

**2. Beta software has a dependency graph, and Windows makes it worse.**
My install chain ended up being: Windows → Node.js → Claude Code → paid Claude account → OpenClaw → Gateway → token → skill → underlying CLI → that CLI's own auth. Each layer is a point of failure. On WSL2 or Linux, most of these are smoother. Native Windows compounds the friction.

If you're on Windows and serious about using OpenClaw, spend the 15 minutes to install WSL2 first. Future you will thank you.

**3. The `SKILL.md` idea is still worth getting excited about.**
Even without finishing the install, reading through the bundled skills taught me something about how to structure *any* agent instruction. The good ones have: a sharp `description` (the trigger), explicit `when to use` rules, concrete examples with expected output, and explicit "don't do this" clauses. That structure is reusable even outside OpenClaw — it's just good prompt engineering discipline, enforced.

I started writing a skill for student inbox triage (URGENT / REVIEW / APPLIED / ARCHIVE buckets for placement-season email). I didn't get to run it. But writing it forced me to articulate what "important email" actually means to me, which is a clarity exercise I don't think I would've done otherwise. That alone was worth the hour.

## Would I keep using it?

Honestly? Probably not right now. I'm a student mid-placement-season. My bandwidth for debugging beta CLI tooling on native Windows is zero. If I come back to OpenClaw, it'll be on WSL2 during a break.

But I'd recommend it to someone with a Mac, spare time, and an inbox they want to tame. The underlying idea is right. The Windows-native onboarding is just rough.

## TL;DR for Windows students

- **Install WSL2 first.** Don't skip the warning.
- **You need a paid Claude plan** (Pro or above). The free tier won't work.
- **Budget 2–3 evenings**, not 1 hour. There are multiple auth flows to get through.
- **Explore `openclaw skills` immediately** — it's the fastest way to understand the model.
- **The `SKILL.md` structure is useful even if you never run the skill.** Treat writing one as a clarity exercise.

Good luck, and happy clawing. 🦞