- Published on
My simple & no-nonsense OpenClaw 🦀 installation journey in $20
- Authors
- Name
- Talha Tahir
- linkedin @thetalhatahir
TL;DR 🎯
This is my journey to a stable OpenClaw 🦀 deployment on a $20 budget, with minimum security risks and minimum headaches. Spoiler: AWS LightSail is a trap, Hostinger's Docker VPS works fine, don't start with Opus, Telegram is banned in Pakistan, and Slack wins the messaging app battle.
If you just want the settings that saved my wallet — scroll to the bottom.
Day 1: AWS LightSail — Looked Perfect, Was Useless 🔴
When I first learned about OpenClaw 🦀, AWS LightSail seemed like the obvious choice. Cloud-native. One-click deployment. Pre-built blueprint. "Secure by default."
All of that is technically true. None of it is useful for OpenClaw 🦀.
What actually happened:
The LightSail OpenClaw 🦀 blueprint is sandboxed in every sense of the word. You can't execute meaningful shell commands. The exec tool throws permission errors. The filesystem is locked down. You can't open ports freely.
But the real dealbreaker? AWS pushes you hard toward Bedrock. It's not that you absolutely cannot use your own Claude API key — it's that reconfiguring the setup away from Bedrock is genuinely painful, undocumented, and fighting against the grain of how LightSail is designed. I wanted to use my own Anthropic API key. That turned into a rabbit hole I never got out of.
After a full day of troubleshooting, the conclusion was clear: LightSail is built for simple stateless web apps. Not for AI agent systems that need shell access, custom services, and API key control. It's a feature, not a limitation — just not a feature I needed.
The damage:
- ⏱️ Full day wasted
- 💸 ~$2 spent on a dead end
- 🧠 Learned: "cloud-native" doesn't mean "agent-friendly"
Day 2: Hostinger KVM-2 VPS — It Works, But There's a Layer 🐳
That evening I found Hostinger's OpenClaw 🦀 VPS option. Tried it the next day. OpenClaw 🦀 was running within 15 minutes.
Before you go all-in though, there's something worth knowing: Hostinger actually offers two separate products here.
- 1-Click OpenClaw 🦀 — fully managed, no root access, Hostinger handles updates and security. Good for beginners who just want things to work without touching a terminal.
- OpenClaw 🦀 on VPS (KVM) — you get a KVM VPS with Docker pre-installed. OpenClaw 🦀 runs inside a Docker container on your server. You have full root access to the VPS itself.
I went with the KVM-2 route. Full server control, which is what I wanted. The tradeoff is that OpenClaw 🦀 lives inside a Docker container, so there's one extra layer of abstraction when you're poking at configs or environment variables. Not a dealbreaker, but worth knowing upfront.
What the setup looked like:
- Spun up Hostinger KVM-2 (~$8-12/month, 2 vCPU, 8GB RAM, 100GB NVMe)
- Docker was pre-configured, OpenClaw 🦀 container was already running
- Full root SSH access — set API keys as environment variables, done. Worth noting: keeping secrets in env vars and out of config files is a security layer in itself 🔒
- 8GB RAM is the sweet spot. Anything less and you'll feel it.
One annoyance: every time the container restarts, you have to enter the gateway token to re-pair your local OpenClaw 🦀 with the VPS instance. It sounds tedious but it's actually a security feature — you're pairing sessions, not blindly trusting the server. I made peace with it.
The $20 Mistake: Starting With Claude Opus 🔥💸
This is the part where I learned an expensive lesson. I got OpenClaw 🦀 running, added my Anthropic API key, and thought: "I'll go with the best model — Opus."
That was a mistake.
Here's what I didn't know: OpenClaw 🦀's initial setup phase is extremely token-heavy. Before you type a single message, it loads system prompts, tool definitions, and configuration scaffolding. You're already burning tokens just getting the session initialized.
Then add the fact that OpenClaw 🦀 sends the full conversation history with every request by default. Every turn, the entire history goes back to the model. With Opus, every one of those input tokens is expensive.
And on a fresh Anthropic Tier 1 account? The real constraint isn't the context window (that's model-level, same across all tiers) — it's the rate limit: Opus on Tier 1 caps you at 30,000 input tokens per minute. OpenClaw 🦀's history-heavy requests hit that ceiling constantly.
Result: $20 burned in less than 48 hours. Just chatting. Nothing intensive.
The fix: Start with a free or cheap model while you're setting up and learning your usage patterns. You don't need Opus for configuration and casual assistant tasks. I eventually landed on gemini-3.1-flash-lite-preview as my primary model — it has a generous free tier and handles instruction-based assistant tasks well. I use claude-haiku-4-5 as a fallback when Gemini limits hit.
The Messaging App Graveyard 📱⚰️
With OpenClaw 🦀 running, I needed a way to actually talk to it through a proper messaging interface. OpenClaw 🦀 supports several integrations. I tried all of them.
Telegram 🚫
Telegram is banned in Pakistan. I knew this going in, but figured I'd create the bot via VPN and use it that way. The bot setup worked fine — I got it connected. But having to fire up a VPN every single time I wanted to message my own AI assistant? That's too much friction. Dropped it.
WhatsApp 🤦♂️
WhatsApp requires you to use your own phone number. There's no way around this. And the experience of messaging yourself to talk to an AI is... deeply weird. You're in your own WhatsApp chat, sending messages to yourself, waiting for a bot to reply. I couldn't get over the psychological weirdness of it. Dropped it too.
Discord ❌
Went through all the hoops — created the bot, configured permissions, connected it to OpenClaw 🦀. 5 hours of my life. It worked. The next day the bot had gone completely stale. Stopped responding. I didn't want to debug a Discord bot on top of everything else. Dropped.
Slack ✅
This is where I landed and where I've stayed.
I already use Slack for work, so the interface felt natural. The setup does take some effort — you need to create a new workspace and create a Slack app with the right bot permissions. But here's the thing: OpenClaw 🦀 actually walks you through the whole process. It tells you exactly which OAuth scopes and authorizations your bot needs. No guessing, no documentation spelunking. The hand-holding made a real difference.
Once set up, it's been rock solid. Messages go in, responses come back. No staleness, no VPN, no weirdness.
Taming the Token Burn — Settings That Actually Work ⚙️🧠
Even after switching to cheaper models, I was still burning through tokens faster than expected. The root problem was OpenClaw 🦀's default behavior: it sends everything — full history, old messages, all of it — on every request. That's fine if you have a budget to throw at it, but my whole goal was to run this as lean as possible.
OpenClaw 🦀 suggested some memory configuration tweaks, and after applying them, things stabilized considerably. Here's exactly what I'm running:
contextTokens: 60000 # max context window size
compactionMode: safeguard # auto-compacts when context gets bloated
maxHistoryShare: 0.6 # max 60% of context can be history
recentTurnsPreserve: 5 # always keep last 5 turns intact
keepLastAssistants: 8 # retain up to 8 assistant messages
The compactionMode: safeguard is the key one. When your context window fills up, OpenClaw 🦀 automatically compacts the history — summarising old turns instead of dropping them entirely. This means you don't lose continuity, but you also don't keep paying for the full verbatim history of every conversation.
Note: These settings work for me and my goal of running lean. If budget isn't a concern for you, stick with the defaults — tweaking these has tradeoffs. Tighter history means the model has less context to work with, which can occasionally affect the quality of responses.
Also worth mentioning: if your primary use case is writing or reviewing code, OpenClaw 🦀 is not the right tool. Claude Code is purpose-built for that. OpenClaw 🦀 shines as a general-purpose virtual assistant — tasks, reminders, research, writing, Q&A. That's exactly how I use it.
My current model setup:
- 🥇 Primary:
gemini-3.1-flash-lite-preview— free tier, fast, great for assistant/instruction tasks - 🔄 Fallback:
claude-haiku-4-5— kicks in when Gemini limits are hit
Here's where I've landed cost-wise:
- 🖥️ Hostinger KVM-2 VPS: ~$12/month — fixed, predictable
- ✨ Gemini Flash: free, as long as I keep my prompts meaningful and don't spam it. I haven't hit the limits yet with normal usage.
- 🦀 Haiku 4-5 fallback: when Gemini limits kick in, Haiku takes over. Running around $0.50/day on Haiku — which is minimal for a personal AI assistant that's available 24/7.
All in, I'm at roughly around 20 USD/month for a personal AI assistant running 24/7. The 30 USD I burned figuring all this out was the real cost — and hopefully this article saves you from paying the same tuition.
What I'd Tell Myself on Day 1 📋
Skip AWS LightSail. It's a sandbox dressed up as a deployment platform. Not built for agent workloads.
Hostinger KVM-2 VPS works. OpenClaw 🦀 runs in Docker on top of it — one extra layer, but fully functional. Full root access, your API keys, your control.
Do not start with Opus. OpenClaw 🦀 is token-hungry by design, especially during setup. Start with Gemini Flash (free) or Haiku. Move to a more powerful model only when you know exactly what you're using it for.
Slack is the best messaging integration. It's the only one where setup is guided, the interface is already familiar, and it doesn't require a VPN, your personal phone number, or debugging a stale bot.
Tune the context settings early.
compactionMode: safeguard, 60k context tokens, 60% max history share or find your sweet spot. Do this before you burn through your budget figuring it out the hard way.