OpenClaw Meets AWS: End-to-End Testing and Deployment

OpenClaw is the most hyped open-source personal AI agent currently being talked about in the community. It allows users to run a fully autonomous assistant. Gone are the days when you just chat with LLMs or configure agents to do predefined work—OpenClaw actually does the work for you. It was only a matter of time before someone built something like this.

Git repo: https://github.com/openclaw/openclaw

There is solid documentation there to help you get started.

OpenClaw is awesome. Let me put it this way: I spent last weekend testing OpenClaw, and here are my key takeaways.

I deployed OpenClaw on AWS EC2.

I listened to people who had tried it before me and didn’t take the risk of deploying it on a personal machine. Instead, I used AWS EC2 to configure and run OpenClaw.

Setup details:

• OS: Amazon Linux
• Instance type: t3.small
• Storage: 30 GB

Everything worked smoothly for the tests I ran. The instance came up without any issues, and OpenClaw operated reliably with no noticeable performance problems.

If you want to install OpenClaw on Linux, it’s incredibly simple—just one command:

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

Connecting WhatsApp to interact with OpenClaw

I initially tried to connect Telegram, but unfortunately my account had limited access and I wasn’t able to create bots. So I went with WhatsApp instead. It was straightforward and painless—probably the easiest approach.

Connecting an LLM with OpenClaw

OpenClaw needs an LLM to work its magic. During the initial configuration, I chose google/gemini-2.5-flash-lite. It’s part of the free tier, and I was able to run a few tests without any issues.

Moving to other LLMs with OpenClaw

Switching the LLM after the initial configuration was a bit tricky, especially when I wanted to connect DeepSeek. I was surprised to see that DeepSeek wasn’t listed in the initial configuration wizard. But no worries—OpenClaw supports the OpenAI standard, and after a few attempts, I was able to configure DeepSeek successfully.

At first, I tried configuring it manually by editing

~/.openclaw/openclaw.json

but later I found an easier approach.

Changing the LLM using the command line
openclaw config set models.providers.deepseek '{
  "baseUrl": "https://api.deepseek.com",
  "apiKey": "<include your API key>",
  "api": "openai-completions",
  "models": [
    { "id": "deepseek-chat", "name": "DeepSeek Chat", "contextWindow": 64000 },
    { "id": "deepseek-reasoner", "name": "DeepSeek R1", "contextWindow": 64000 }
  ]
}' --json

Operating OpenClaw using the terminal

To launch the terminal UI, it’s just one command:

openclaw tui

Online search capability with the Brave Search API

I wanted OpenClaw to have online search capabilities, so I used the Brave Search API. It’s free and comes with a generous free tier.

You’re ready to go.

That’s it—OpenClaw connected to my WhatsApp and started doing the magic for me.

Issues I observed

From time to time, OpenClaw would hang and return NO_REPL.

NO_REPL usually means the agent or session is not running in an interactive command REPL. Instead, it’s operating in a managed or controlled mode.
The bottom line: I was connected, but not dropped into a live command shell.

When I stopped getting responses, I did what anyone would do—restarted the EC2 instance.

Occasionally, I realized it was stuck in the terminal but still responding on WhatsApp. Since OpenClaw was working well for me at that point, I didn’t dig into it further.

Did the agent do anything that suggested it might go rogue?

Yes—slightly.

I gave the agent access to one of my GitHub repositories, and it dumped not only the files we were working on for the project, but some others as well. I think OpenClaw thought it was a nice dumping site 😄

I didn’t investigate this further, but this is the only instance where I noticed that behavior.

Use case I tried

Complete website development—from development to deployment.

Let me share the steps I followed:

• I wanted to develop a single-page website.
• I looked for a template online.
• I gave the template to the agent and asked it to build something similar.
• It ended up being more of a white-label site.
• I provided my requirements document link, and the agent was able to complete the site.
• Obtaining images was challenging since it only had API-based search access.
• Still, it managed to pull some decent images that made the site look good.
• I gave the agent access to GitHub, and it pushed the code to the repository as well.
• Working on changes was easy—it didn’t complain or resist updates.
• There were only very minor bugs; only twice did it fail to bring up the site.
• It successfully handled the full deployment process too.

Overall, my test proves that this agent is capable of handling end-to-end software development with minimal human guidance, while the agent does most of the heavy lifting.

OpenClaw Troubleshooting: Issues & Solutions (Ongoing Guide)

Issue: WhatsApp Stops Responding (Even though OpenClaw is running)

The server shows as active in the terminal, but the bot isn't replying to messages on WhatsApp.

The Fix:
Instead of manual debugging, I asked the Agent to check the connection; it diagnosed the "silent" session and autonomously triggered a refresh of the WhatsApp handshake. It fixed its own connectivity in the background without me typing a single restart command—true self-healing AI.