March 9, 2026

What is OpenClaw? The Complete Guide (2026)

OpenClaw is an open-source AI agent that runs on your machine and takes real actions via WhatsApp, Telegram, WeChat, and more. Here's everything you need to know.

AdSense · 728×90

In late 2025, a GitHub repository crossed 247,000 stars in under two months — faster than almost anything before it. The project was called OpenClaw, and it did something the AI world had been waiting for: it let you message an AI agent on your phone and have it actually do things for you. Not simulate doing things. Not generate a plan. Actually execute.

This guide covers everything you need to know about OpenClaw — what it is, how it works, why it went viral, and whether it’s right for you.

What Is OpenClaw?

OpenClaw is an open-source AI agent framework that runs as a persistent background process on your local machine. Unlike chatbots that live in a browser tab, OpenClaw integrates directly with the messaging apps you already use — WhatsApp, Telegram, WeChat, Slack, and more — so you can send it commands the same way you’d text a friend.

The defining characteristic of OpenClaw is that it takes real actions:

  • Email — read, draft, send, archive, label
  • Calendar — create events, reschedule meetings, send invites
  • Web browsing — search, summarize pages, fill forms, extract data
  • Shell commands — run scripts, manage files, automate repetitive tasks
  • Code execution — write and run Python, Node, or shell code inline
  • Negotiation flows — multi-step tasks that require back-and-forth with external services

Everything runs on your hardware. Your messages never touch OpenClaw’s servers. The AI model call goes to whichever provider you configure (Anthropic, OpenAI, DeepSeek, Kimi, etc.), but the orchestration layer — the part that decides what tools to call and in what order — lives entirely on your machine.

Who Made OpenClaw?

OpenClaw was created by Peter Steinberger, the founder of PSPDFKit (now Nutrient), one of the most widely used PDF rendering SDKs in the world. Steinberger had been building developer tools for over a decade when he started experimenting with agentic AI systems in early 2025.

The project began as an internal tool called Clawdbot — a glorified macro runner that could handle his email triage while he was traveling. He open-sourced it in September 2025 as OpenClaw with minimal fanfare. Within 72 hours, it had 12,000 GitHub stars.

Steinberger joined OpenAI in February 2026 to lead applied agent research, handing day-to-day stewardship of OpenClaw to the newly formed OpenClaw Foundation, a nonprofit focused on keeping the project open, secure, and community-driven. The Foundation currently has seven full-time maintainers and a contributor base of over 2,400 developers.

Why Did OpenClaw Go Viral?

Two words: Moltbook demo.

In October 2025, a developer named @moltbook posted a 47-second screen recording to X (Twitter). In the video, he sends a single WhatsApp message: “Schedule a meeting with Sarah for next Tuesday, find her a parking spot near the office, and email her the details.”

OpenClaw completes all three tasks in 38 seconds — checking his calendar, running a Google Maps search for nearby parking, drafting the email, and sending it — while he watches from his phone.

The clip hit 14 million views in 48 hours.

What made it land was not the technology itself — AI had been capable of these individual tasks for years. It was the delivery mechanism. You didn’t need a browser. You didn’t need to open a new app. You sent a WhatsApp message from the same place you text your family, and your computer handled the rest.

Within 72 hours of the video going viral, the OpenClaw repository saw over 32,000 new agent deployments logged by the telemetry opt-in. People weren’t just starring the repo — they were installing and running it.

The core insight the project validated: people don’t want AI that talks about doing things. They want AI that does things, delivered through interfaces they already trust.

How Does OpenClaw Work?

At its core, OpenClaw is a gateway process — a lightweight server that runs in the background on your machine and bridges messaging apps to AI models to local tools.

Here’s the flow for a typical command:

  1. Message received — You send a message via WhatsApp, Telegram, or another connected channel. OpenClaw’s gateway picks it up through the platform’s API or an official bot account.
  2. Context assembled — The gateway pulls in relevant context: your SOUL.md persona file, any active Skills, recent conversation history, and current system state.
  3. AI model called — The assembled prompt is sent to your configured AI model (Claude, GPT-4o, DeepSeek-R2, Kimi-2.5, etc.).
  4. Tool calls executed — The model responds with structured tool calls. OpenClaw executes them locally: running shell commands, calling APIs, reading files, opening browsers.
  5. Results reported back — The results are summarized and sent back to you as a reply in the same messaging thread.

Your data stays local. The only outbound network calls are (a) the AI API call with your prompt and (b) any tool calls that inherently reach the internet (like web search or sending email).

OpenClaw uses a SKILL.md system for extending capabilities and a SOUL.md file for defining the agent’s persona and behavioral rules. More on both below.

What Are Skills?

Skills are the primary extension mechanism in OpenClaw. Each skill is a folder containing a SKILL.md file — a plain Markdown document that instructs the AI model how to perform a specific category of tasks.

Skills can also include helper scripts (Python, Node, shell) that the agent calls when needed, but many skills are pure Markdown — just clear English instructions that tell the model what steps to follow.

The community-run ClawHub registry currently hosts over 5,400 skills, ranging from the practical:

  • gmail — full email read/write/label/search
  • calendar — Google Calendar and Outlook integration
  • github — PR review, issue triage, commit summaries
  • obsidian — read and write to your Obsidian vault
  • cron-tasks — schedule recurring agent tasks

…to the creative and niche:

  • wine-pairing — recommend wines based on what’s in your fridge
  • lease-reviewer — parse and flag clauses in rental agreements
  • podcast-digest — summarize your unplayed podcast backlog

Installing a skill from ClawHub takes one command:

clawhub install gmail

You can also drop a skill folder directly into your ~/.openclaw/skills/ directory and it becomes available immediately.

What Is QClaw?

QClaw is Tencent’s officially supported distribution of OpenClaw, built specifically for the Chinese market. It entered internal beta on March 9, 2026 and opened public macOS downloads on March 10, 2026.

Where raw OpenClaw requires a terminal, a config file, and an API key to get started, QClaw wraps the entire setup in a polished desktop application. You download it from claw.guanjia.qq.com (Tencent PC Manager’s domain), click through a short wizard, scan a WeChat QR code with your phone, and your agent is live — typically in under three minutes.

Key differences from upstream OpenClaw:

  • WeChat and QQ integration out of the box, with no manual bot setup
  • Kimi-2.5 as the default AI model (free during preview, no API key required)
  • QClaw Skills Store — a curated subset of ClawHub optimized for Chinese apps and services
  • Built and maintained by the Tencent PC Manager team, with a qq.com domain and official support channels

QClaw is the recommended path for anyone in mainland China, or anyone who wants to use OpenClaw primarily through WeChat or QQ.

Is OpenClaw Safe?

OpenClaw is open source, actively audited, and designed with a local-first architecture that minimizes data exposure. That said, it is a system that executes code and performs real actions on your behalf, which introduces genuine risks worth understanding.

Cisco Talos published a report in January 2026 flagging prompt injection as the primary threat vector. If OpenClaw is instructed to browse a malicious web page, that page could contain hidden instructions designed to hijack the agent’s next action — for example, telling it to forward your emails to an attacker-controlled address.

Practical safety guidelines:

  • Only install skills from trusted sources. The ClawHub top-100 list is a reasonable starting point. Read the SKILL.md before installing anything from unknown authors.
  • Restrict shell access. In config.yaml, set allow_shell: false if you don’t need that capability. Most common tasks don’t require it.
  • Run locally only. Don’t expose OpenClaw’s local port to the internet unless you know exactly what you’re doing.
  • Update regularly. The Foundation patches security issues quickly. Run openclaw update weekly.
  • Review agent logs. OpenClaw keeps a full action log at ~/.openclaw/logs/. Glance at it periodically.

The local-first design is itself a safety feature — your emails, files, and calendar data are processed on your machine, not sent to a third-party agent cloud. Compare this to cloud-based agent services where all your connected data passes through someone else’s infrastructure.

How Do I Get Started?

Path 1: QClaw (easiest)

If you’re on Windows or Mac and want WeChat or QQ integration, download QClaw from claw.guanjia.qq.com. You’ll be up and running in under five minutes with no terminal required.

As of March 10, 2026, official downloads are available for macOS (Apple Silicon + Intel). Windows is still listed as coming soon.

Path 2: Raw OpenClaw (most flexible)

If you want full control, Telegram/WhatsApp integration, or you’re running Linux, install OpenClaw directly:

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

Then configure your AI model in ~/.openclaw/config.yaml, connect a messaging channel, and install your first skills from ClawHub.

Both paths lead to the same underlying agent. QClaw is a launcher that handles the plumbing. Raw OpenClaw gives you direct access to every configuration option.

AdSense · 336×280