I gave OpenClaw its own computer and a Telegram Bot. Here's What Happened

Quick thing before we get into it: if you're thinking about running OpenClaw yourself, skip putting it on your home or work machine for security reasons, of course. Get a cheap VPS.

I used AWS because I work in the cloud daily and know how to keep costs in check, but for most people, especially if you've never managed cloud infrastructure before, AWS bills can sneak up on you fast. That's why I'd point non-cloud users straight to Hostinger Open Claw VPS.

Their VPS plans are simple, predictably priced, and you won't wake up to a surprise bill because you forgot to set a budget alert.

Okay. Now the story.

What OpenClaw Actually Is (Because Most People Get This Wrong)

It's not a chatbot. It's not a SaaS dashboard. It's not another wrapper around GPT with a monthly subscription slapped on it.

OpenClaw is an open-source personal AI assistant that runs on your own machine or server, connects to whichever AI model you choose, and talks to you through whatever chat app you already use.

Telegram, WhatsApp, Discord, iMessage, Slack. You message it like a contact. It does things.

Because it runs on your infrastructure and not someone else's servers, your data stays yours.

You pick the model from a list of various top models like DeepSeek Claude, GPT-5.2, or local models if you want. You control what it can access. And since it's fully open source, you can extend it however you need to.

OpenClaw was built by Peter Steinberger and a growing community of contributors. It's only a few months old and moves fast — new capabilities show up regularly.

My Setup

Ubuntu 24.04 VPS on AWS Lightsail costs $12/per month. Mistral, as the AI backbone, routed through a Vercel API gateway I set up as an abstraction layer, makes it easier to swap models later without touching every config file. Telegram for delivery.

I configured it to send me a structured morning report at 7:00 AM East Africa Time every day. For the last week, it has worked exactly as intended. Getting there, though, was a different story.

The First Thing I Got Wrong: Claude Opus 4.6 Ate My Credits

When I first installed OpenClaw, I defaulted to Claude Opus 4.6 as the model. Makes sense in theory, best output, most capable, why not start with the best?

My credit balance had other opinions.

Opus 4.6 is an incredible model. It's also priced like one. For a daily scheduled task firing the same structured request every morning, using Opus is like going swimming in the local pool with full scuba diving equipment.

What is the output quality difference compared to a cheaper model on this kind of repetitive, well-defined task? Honestly negligible. The price difference? Not negligible at all.

I switched to Mistral after the first few hours and haven't looked back. The reports are just as useful. The bill is much easier to live with.

Lesson: match the model to the task. Opus-level models are worth it for complex, nuanced, one-off reasoning. For structured daily reports pulling from consistent data, use something cheaper and cache what you can.

On Caching: Seriously, Do This

If you're running OpenClaw for repetitive scheduled tasks, look into caching your API calls.

Any part of your prompt that doesn't change between runs, system instructions, static context, or your data schema doesn't need to be re-sent and re-processed every single time.

Most major AI APIs support prompt caching in some form, and the cost savings on high-frequency tasks are real.

The general principle:

• ❌ Expensive model + no caching + daily repetitive task = you will regret it
• ✅ Cheaper model + cached static context + well-structured prompt = same job, fraction of the cost

Why I'm on AWS But Suggesting Hostinger for Most People

I work in the cloud professionally. Setting up billing alerts, IAM permissions, configuring CloudWatch dashboards, monitoring resource usage, and watching for cost anomalies is just part of how I operate every day. So AWS Lightsail made sense for me.

I knew how to keep it under control.

I’ve watched people with no cloud background spin up resources, forget they even exist, and then get blindsided by a massive bill weeks later.

A quick scroll through the Google Cloud subreddit is all it takes. The horror stories from mismanaged infrastructure are a reminder that in the cloud, what you forget can cost you

AWS is powerful precisely because it has so many knobs to turn, and if you don't know which ones matter, that flexibility becomes a liability.

For anyone who doesn't live in cloud infrastructure, I'd genuinely recommend starting with Hostinger's Openclaw VPS plans.

Clean Ubuntu setup, predictable flat-rate pricing, no need to understand reserved instances, Route 53 or any of the things that make AWS bills confusing. You get a server, it costs what it says it costs, you SSH in and get to work.

A basic 2 GB RAM instance is more than enough for what OpenClaw needs, but Hostinger is even more generous and gives you an 8GB RAM 2vCPU and 8TB bandwidth at $7.99/mon

If you are already familiar with AWS or GCP, use what you're comfortable with.

If you don't, don't learn cloud infrastructure and OpenClaw at the same time. That's just unnecessary pain.

Days 1 and 2: The Part Nobody Writes About

Every write-up about a new tool skips straight to "and then it worked great." So let me be the one to actually tell you what the first 48 hours looked like.

The cron job that was definitely running (it wasn't)

My plan: schedule the morning report with a cron job. Simple. Except nothing happened. No error. No log. No indication that anything had been attempted. Just silence.

Turned out to be three things piling on each other simultaneously:

1. Environment variables set in my normal shell session weren't available to the cron environment
2. The virtual environment path wasn't being activated correctly inside the cron command
3. I'd set the schedule for 7 AM without accounting for the server running UTC, not EAT. I needed 04:00 UTC, not 07:00

None of these is hard to fix once you know what's wrong. Finding all three without clear error messages at 11 PM was a different experience.

Always add logging to your cron jobs from day one. Redirect output to a file. Silent failures are the worst kind.

your-command >> /home/user/openclaw/cron.log 2>&1

Manual requests: educational, but a grind

While the cron issue was unresolved, I was testing by firing requests manually from the terminal. Run the command. Watch the output. Tweak something. Run again.

It works as a debugging method. As a workflow, it's a bumper, especially when you're used to just typing in Telegram and having things happen.

The upside: by the time the scheduler was working, I understood every layer of the system. The manual phase taught me OpenClaw better than any documentation would have.

Day 3, 7:02 AM

Telegram notification. Report there. Automated. No manual intervention.

The relief was disproportionate to the actual achievement, but completely understandable given the previous 48 hours. It's been delivering every morning since.

What Else People Are Doing With This Thing

My scheduled report is a pretty narrow use case. Looking at what the broader community has built:

• Email triage and sending, calendar management — all from a WhatsApp or Telegram message
• Remote coding sessions via Claude Code or OpenAI Codex — kick off a task from your phone, get results in chat
• Image generation through Nano Banana — request visuals in chat, receive them without switching apps
• Browser control for automated form filling, flight check-ins, and data extraction from sites without APIs
• Health and biometric tracking — WHOOP, Oura, and other wearables connected and queryable in chat
• Obsidian and personal knowledge base integration — builds your second brain from conversations
• Smart home control — lights, air purifiers, environmental conditions managed against defined goals
• Custom audio content — one person had it write meditation scripts, convert to TTS, add ambient sound, and deliver the finished file

Check out open claw subreddits for ideas of tasks you can automate with openclaw

The pattern across all of these is the same: friction that used to require opening multiple apps gets compressed into one message in a chat thread you already have open.

That's the value proposition, and once you feel it working, it's hard to go back.

Honestly? I'm Kind of Hooked.

I didn't expect to feel this way about infrastructure software. But there's something genuinely satisfying about a system that works while you're not working, the report arrives before I've opened a browser, before the noise of the day starts.

Information I specifically configured to matter to me, already processed, already waiting.

And the thing about OpenClaw is that you start building around it almost immediately. Within a few days, I was already thinking about the next automation, the next friction point that a scheduled task could eliminate.

That's a different relationship than most software creates.

What's coming next for my setup

I'm adding the Gemini API alongside Mistral, specifically Gemini Nano for lightweight, fast tasks that don't need heavy models. Paired with Nano Banana for image generation, I can start building visual content workflows directly from Telegram without switching tools.

The one I'm most excited about: OpenAI API with Codex access. Triggering remote coding sessions from my phone, having OpenClaw run code, fix tests, open PRs, and report back while I'm doing something else is exactly the kind of async workflow this thing is built for. Community members are already doing this, and the demos I see on Discord are wild.

Beyond specific APIs: a broader daily automation layer. Not just one morning report, but a set of scheduled and on-demand tasks that handle the routine stuff, email triage, content summarisation, task reminders with context, all through Telegram. Fewer context switches, not more tools.

Should You Try This?

If you're comfortable with a terminal and patient enough to debug environment issues, yes. The first two days will test you. Day three makes it worth it.

A few things that'll save you time:

• Don't start with the most expensive model. Mistral Devstral or Qwen3 is fine for most scheduled tasks. Move up when you actually need to.
• Cache your static prompt content from the start. Your system instructions aren't changing between runs — don't pay to send them every time.
• Match model to task complexity. Daily structured reports don't need Opus or GPT 5.3. Complex one-off reasoning might. Know the difference.
• Add logging to cron jobs immediately. Future you will be grateful.
• Set UTC times for everything server-side. Convert locally.

If you want to run this without cloud billing anxiety, Hostinger's VPS plans are the low-drama option —flat pricing, Ubuntu setup in under 20 minutes, no surprise bills. Once the server is live, everything else follows from the OpenClaw docs.

Set aside two to three days for initial setup. Give yourself permission to find the first 48 hours frustrating.

Day three is worth it.