There is a new kind of AI tool making waves. It is called OpenClaw, and it acts and does things.
OpenClaw is an open-source, self-hosted AI assistant created by Peter Steinberger. Launched as "Clawdbot" in November 2025, it is now called OpenClaw and has 145,000 GitHub stars. Nature, Wired, DigitalOcean, and VirusTotal have covered it.
But what exactly is it? And more importantly, is it something you should set up?
Is OpenClaw for Everyone? No.
OpenClaw is powerful, but it is not a plug-and-play consumer app. One of OpenClaw's own maintainers put it bluntly: "If you can't understand how to run a command line, this is far too dangerous a project for you to use safely."
Here is why. OpenClaw runs on your machine with access to your files, shell, browser, and messaging apps. It installs extensions called "Skills." In February 2026, security researchers found over 300 malicious skills on ClawHub (the skill marketplace), disguising malware as helpful tools. A Cornell audit found 26% of skill packages contained vulnerabilities.
OpenClaw partnered with VirusTotal for automated skill scanning and hired a security advisor. But they are clear: this is not a silver bullet. Prompt injection attacks—where malicious instructions hide in natural language—remain unsolved.
If you understand what's installed on your system, can audit code, and know command-line tools, OpenClaw is exciting. If not, wait until it matures or work with someone who can set it up for you.
How does OpenClaw Work—Through the Gateway?
At OpenClaw's core sits the Gateway—a Node.js process that controls everything.
Think of it as an Agent Server. The Gateway is responsible for managing all your channel connections (WhatsApp, Telegram, Slack, Discord, and more), routing messages to the right AI agent, maintaining session state and conversation history, coordinating tool execution (file operations, shell commands, browser automation), and providing a WebSocket interface for CLI tools, mobile apps, and the web UI.
Every message, whether from WhatsApp or Telegram, flows through the Gateway, gets routed to the right agent, and the response flows back. By default, it runs on localhost:18789, keeping everything local.
What makes this interesting: you can run multiple agents behind one Gateway, each with its own personality, workspace, model, and tool access. For example, route WhatsApp messages to Claude Sonnet and Telegram messages to Claude Opus, all from one Gateway.
It is microservices for personal AI.
For the Masses, What Does OpenClaw Actually Do?
OpenClaw is an AI assistant that performs actions. Unlike ChatGPT or Claude's chat interface, which respond with text, OpenClaw connects to your tools and systems. It reads and writes files on your computer, executes shell commands, browses the web, sends messages through your apps, and interacts with APIs.
This is organized through:
Agents: The AI brains with isolated workspaces and sessions
Tools: Built-in capabilities like exec, read, browser, cron, and message
Skills: Community or custom packages that teach the agent how to use specific tools and APIs, like GitHub, Notion, Slack, or Google Calendar
In plain language: You tell it what to do—through any messaging app you use—and it does it.
How Do You Talk to Your OpenClaw AI Assistant?
You do not need a special app to create an AI assistant with OpenClaw. This is one of OpenClaw's best design decisions. You use the messaging platforms you already have.
OpenClaw supports WhatsApp, Telegram (recommended for first setup), Slack, Discord, Signal, iMessage, Google Chat, Microsoft Teams, and WebChat. It also supports Matrix, BlueBubbles, Zalo, and more. It can speak and listen on macOS, iOS, and Android through companion apps.
The experience: message your bot on Telegram, and it responds. It does not respond with just words but with actions as well.
Can OpenClaw Schedule Things for You?
Yes. And it is built right in.
OpenClaw includes a powerful built-in scheduler called Cron, which runs inside the Gateway process itself. It supports full 5-field cron expressions with timezone awareness, one-shot reminders (e.g., "remind me at 4 PM"), recurring jobs (e.g., "every weekday at 9 AM, summarize my unread emails"), isolated agent turns where a scheduled job runs in its own session, and delivery to any connected channel — so the result can show up in your WhatsApp, Telegram, or Slack.
Jobs persist under ~/.openclaw/cron/, so even if the Gateway restarts, your schedules survive.
For example, you could set up a job that checks your email every 10 minutes. If it finds a message from a specific sender — say your manager or a key client — it immediately notifies you on WhatsApp. People are doing this today.
What About Webhooks and External Triggers?
OpenClaw's Gateway provides a WebSocket interface and can be integrated with external systems in multiple ways. You can trigger agents via CLI commands, connect through the WebSocket API, set up cron-based polling workflows, and use skills that integrate with external services and APIs. The architecture is deliberately open-ended, encouraging creative integrations.
Real Use Cases that People are Building
Here is where things get exciting. OpenClaw is a productivity multiplier. People are using it to manage their calendars by scheduling events, getting daily briefings, and rescheduling conflicts through a simple chat message. They are processing emails, with agents that scan your inbox, summarize important threads, draft replies, and flag urgent messages from specific senders.
Developers are running coding agents in the background, using skills for Claude Code, Codex, and other code-generation tools that build features, run tests, and open pull requests — while they sleep. One user described kicking off Claude Code sessions from his phone and waking up to completed PRs.
Early adopters are automating household workflows. One user configured OpenClaw to build a weekly meal planning system in Notion, saving his family an hour every week. Others are monitoring health data, connecting WHOOP and other services to track biomarkers and daily habits through chat.
Some have even built entire websites from their phones, controlling file operations and deployment through Telegram messages.
The pattern is that anything you would ask a skilled executive assistant to do. The assistants check information, draft content, and provide reminders. Anything you need to keep an eye on, it can be handled by OpenClaw.
Example of OpenClaw Usage
Wednesday, 6 AM: Your assistant checks email. Client needs a deliverable update. It drafts a reply and waits for your approval. Your accountant sent something about taxes - flagged as urgent. The newsletter you follow gets summarized and saved to Notion.
7 AM: Health check on your staging server. Green. No alert.
8 AM: Morning briefing hits your phone. "Three meetings today. First one at 10 AM. Two-hour focus block this afternoon. Your mom's birthday is tomorrow. Want me to order flowers?"
You are still in bed. You type: "Yes."
That is OpenClaw, running right now, on someone's laptop.
Can I run it Completely On-Premise?
Yes. This is a key differentiator from existing chatbot or assistant AI builders.
OpenClaw is model-agnostic. While it works beautifully with cloud-based models like Claude, GPT-4, and DeepSeek, you can run it entirely on local models using Ollama. This means zero data leaves your machine.
Supported local models include Llama 3.x, Mistral, Qwen, Phi, DeepSeek-R1, and more. The setup is straightforward: install Ollama, pull a model, point OpenClaw's config at your local endpoint (http://127.0.0.1:11434/v1), and you're running a fully private, on-premise AI assistant.
OpenClaw even supports hybrid configurations — use a fast local model as your everyday coordinator and fall back to a cloud model for complex reasoning tasks.
For organizations with strict data residency or compliance requirements, this is a game-changer.
A Word on Security — Take It Seriously
OpenClaw gives an AI agent real access to your system. That power comes with real risk.
In the past few weeks, researchers discovered 341 malicious skills in a campaign called "ClawHavoc" that targeted cryptocurrency wallets. Cisco's security team found third-party skills performing data exfiltration without user awareness. BitSight and CrowdStrike flagged exposed instances as enterprise security risks. 22% of enterprise customers in one study had employees running OpenClaw without IT approval.
OpenClaw is improving: VirusTotal partnership, sandbox modes for Docker-isolated execution, tool allowlists and denylists, DM pairing for access control, and a formal security program at trust.openclaw.ai.
But the responsibility falls on you:
Audit skills before installing them
Enable sandbox mode
Restrict tool access per agent
Keep sensitive credentials out of the agent's reach
If deploying for a team, treat it like privileged infrastructure—with governance, monitoring, and clear boundaries
The Mental Model Shift
The tool is only as useful as the workflows you design for it.
Setting it up is just the beginning. The real work—and the real value—comes from thinking about what you want automated. What tasks do you repeat every day? What decisions could be offloaded? What monitoring would save you from fire drills?
OpenClaw is not magic. It is infrastructure. And like any infrastructure, it rewards planning.
You need a mental model for what is possible, then build, iterate, and refine. The people getting the most out of OpenClaw are not the ones with the fanciest hardware—they're the ones who've thought about their workflows.
At GeekyAnts, we help people plan, build, deploy, and manage personal AI assistants—from initial architecture to production automation. Whether you want to run OpenClaw on-premise, integrate it with your existing tools, or build custom skills for your team, reach out.
This article was originally published on the GeekyAnts blog.
