This is a submission for the OpenClaw Writing Challenge
I'm going to be straight with you: I did not finish my OpenClaw project today.
I set out to build ClearView — an accessibility skill that would let blind and low-vision users describe their screen, read documents, and navigate their computer just by messaging a Telegram bot. Great idea. Real use case. I was excited.
What I actually spent my day doing was fighting Windows.
And I'm writing this post because when I was googling every error I hit, I found nothing useful. Every OpenClaw tutorial assumes you're on a Mac. Every setup guide breezes past the "install it and run it" step like it's nothing. For Windows users — especially those using nvm4w — it is very much not nothing.
So here's the guide I wish existed. The honest one. With the real errors and the real fixes.
What I Was Building
The idea was simple: a skill for blind and low-vision users that lets them message their bot and get back a plain-English description of whatever is on their screen, or have documents and images read aloud to them. No new apps to learn. No complex interface. Just Telegram, OpenClaw, and Claude's vision capability.
It's a real problem worth solving. Screen readers are expensive, rigid, and require training. Claude can describe a screenshot in plain language instantly. The gap between those two things is exactly what OpenClaw is built to bridge.
I didn't finish the skill. But I did get OpenClaw running on Windows — and that alone took most of the day. Here's everything I learned.
The Setup: What the Docs Say vs What Actually Happens
1. Node version matters — but so does how you installed it
The docs say you need Node 22 or 24. Fine. What they don't tell you is that if you installed Node via nvm4w (Node Version Manager for Windows), OpenClaw may not find it correctly.
When I first tried to launch the gateway, I got the following:
Error: failed to launch TUI: spawn C:\Program Files\nodejs\node.exe ENOENT
The problem: nvm4w stores Node at C:\nvm4w\nodejs\node.exe, not the default path OpenClaw expects. The fix is to run Command Prompt as Administrator — this ensures the path resolves correctly. Every single time. Don't skip this step.
2. The Telegram plugin EPERM error
When I first ran the gateway, I saw this buried in the logs:
[plugins] telegram failed during register: Error: EPERM: operation not permitted,
rename '....plugin-telegram-lubdRt\plugin' -> '...\telegram'
This is Windows blocking OpenClaw from renaming a temp folder during plugin setup. The fix is the same: run as Administrator. Once I did that, the error disappeared and Telegram loaded cleanly.
3. Setting your Telegram token — the right command
During onboarding I tried:
openclaw config set telegram.token "YOUR_TOKEN"
This gives you:
Error: Config validation failed: : Unrecognized key: "telegram"
The correct command is:
openclaw channels add --channel telegram --token "YOUR_TOKEN_HERE"
That's it. One line. Not in the getting started docs anywhere that I could find.
4. Setting your Claude API key
Similarly, trying to set the Anthropic key through config commands doesn't work intuitively. The way that actually worked for me was:
openclaw configure
Select Model from the sections list, choose Anthropic, and paste your sk-ant- key when prompted.
5. Starting the gateway — use --force
Plain openclaw gateway would start, reach the ready state, then crash with:
Unhandled promise rejection: CIAO PROBING CANCELLED
This is a bonjour/mDNS issue on Windows. The fix:
openclaw gateway run --force
This kills anything stuck on port 18789 first and starts clean. After running this as Administrator, my gateway finally stayed alive.
The Correct Setup Sequence for Windows
If you're on Windows and want to save yourself a day, here is the exact order of commands that works:
1. Open Command Prompt as Administrator
2. Install OpenClaw
npm install -g openclaw@latest
3. Add your Telegram bot token
openclaw channels add --channel telegram --token "YOUR_TELEGRAM_BOT_TOKEN"
(Get your token from @botfather on Telegram — send it /newbot and follow the steps)
4. Set your Claude API key
openclaw configure
Select Model → Anthropic → paste your sk-ant- key
5. Start the gateway
openclaw gateway run --force
6. Wait for this line:
[gateway] ready
That's it. You're running.
What I Actually Learned About OpenClaw
Here's the thing — even through all the frustration, I came away genuinely impressed by what OpenClaw is trying to do.
The skill system is clever. A SKILL.md file with plain-English instructions that the AI interprets at runtime — no complex APIs, no schema, just natural language. That's the right abstraction for this kind of tool. It means anyone can build a skill, not just developers.
The channel architecture is smart too. The same agent, the same skills, accessible through Telegram, WhatsApp, Discord — whatever messaging app you're already using. That's meaningful for accessibility specifically: you're not asking someone with a disability to learn a new interface. You're meeting them where they already are.
The Windows experience is rough right now. The docs assume Unix. The errors are cryptic. The setup is slower than it should be. But the core idea — a personal AI that lives on your own machine, uses your own API keys, and does exactly what you tell it — is genuinely different from everything else out there.
What ClearView Would Look Like (When I Finish It)
For anyone who wants to build the accessibility skill I was attempting, here's the SKILL.md I wrote — it just needs a working gateway underneath it:
---
name: clearview
description: Accessibility assistant. Describes images, reads text, narrates screen content for blind and low-vision users.
---
## ClearView
When the user sends an image or photo:
1. Describe everything visible in plain simple English
2. Read ALL text visible — buttons, menus, labels, errors
3. Describe layout — top, middle, bottom, left, right
4. Describe what buttons and icons do, not just what they look like
5. End with a one sentence summary
When the user sends a photo of a document, letter, or printed text:
1. Read all text from top to bottom
2. Note headings and structure
3. Summarise what it is about
When asked "what's on my screen?":
1. Take a screenshot
2. Describe it fully as above
Rules:
- Say "The screen shows..." not "I can see..."
- Always prioritise text over design descriptions
- If blurry or unclear, say so honestly
Drop that in ~/.openclaw/workspace/skills/clearview/SKILL.md and restart your gateway. That's the whole skill.
The Honest Take
Most OpenClaw posts you'll read this week are going to be polished success stories. Person builds thing, thing works, here are the screenshots.
This isn't that. I spent a full day on setup, hit six different errors, and didn't finish the project I set out to build.
But I think that's worth writing about too — because the Windows experience is a real gap, and the next person who hits CIAO PROBING CANCELLED at midnight deserves to find an answer.
OpenClaw has a lot of potential. The Windows setup needs work. Both things are true.
I'll finish ClearView. Just not today.
Have you built something for accessibility with OpenClaw? Or hit different errors on Windows? Drop them in the comments — I want to know what I missed.
Top comments (0)