The Weekend Project That Dethroned React — and What It Means for Everyone

On March 3, 2026, a four-month-old open-source project did what no software had done in over a decade: it surpassed React as GitHub's most-starred repository. OpenClaw hit 250,829 stars, shattering a record React had built over ten years — in just 60 days of explosive growth. With 48,274 forks, 1,075 contributors, and a skill marketplace hosting over 13,000 community-built plugins, OpenClaw isn't just a trending repo. It's the clearest signal yet that the AI industry is shifting from chatbots to autonomous agents.

What makes this story even more remarkable is its origin. In November 2025, Austrian developer Peter Steinberger spent roughly one hour wiring an LLM to WhatsApp's API, giving it access to web search and file operations. The result — originally called Clawdbot — was an AI that could receive messages, choose its own tools, execute multi-step tasks, and report back through a messaging app. That weekend prototype became the fastest-growing repository in GitHub history, gaining 190,000 stars in its first 14 days.

The OpenAI Acquisition: What Changed and What Didn't

On February 15, 2026, Sam Altman announced that OpenAI had acquired OpenClaw and that Steinberger was joining the company. VentureBeat called it "the beginning of the end of the ChatGPT era" — a declaration that OpenAI itself sees the future in agents, not chat interfaces.

Technically, this was an acqui-hire rather than a traditional acquisition. No price was disclosed. The deal's defining feature was Steinberger's non-negotiable condition: OpenClaw must remain open source. OpenAI agreed, and the project moved to an independent foundation under the MIT license.

Here's what concretely changed post-acquisition:

• Financial backing: OpenAI now sponsors the project and guarantees Steinberger's time for maintenance
• Deeper model integration: Native support for OpenAI's latest models improved significantly
• Foundation governance: Development direction remains community-driven under the new foundation
• Open source preserved: MIT license, full source access, no restrictions

Steinberger's stated mission at OpenAI: "Build an agent that even my mum can use." That aspiration — making autonomous AI accessible to non-technical users — is precisely why OpenClaw resonated with millions in the first place.

Getting Started: Installation in Under 5 Minutes

System Requirements

• Node.js 24 (recommended) or Node 22 LTS (22.16+)
• macOS, Linux, or Windows (WSL2 recommended)

The Fastest Path

On macOS, Linux, or WSL2, a single command handles everything — Node detection, dependency installation, and setup:

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

For Windows PowerShell:

iwr -useb https://openclaw.ai/install.ps1 | iex

If you already have Node installed:

npm install -g openclaw@latest
openclaw onboard --install-daemon

The openclaw onboard wizard walks you through gateway connection, channel configuration, skills, and workspace setup. After installation, three diagnostic commands are your best friends: openclaw doctor (configuration check), openclaw status (gateway health), and openclaw dashboard (web UI).

API Key Configuration

Installation creates a .openclaw/config.yaml file in your home directory. This is where you configure your model provider and API key. A note for Claude users: since Anthropic shut down OAuth access in January 2026, the only connection method is a pay-as-you-go Anthropic API key. OpenAI models connect through their standard API keys.

Docker, Podman, Nix, and Ansible deployment options are documented separately for those who prefer containerized setups.

Connecting Your Messaging Apps: WhatsApp and Telegram

OpenClaw's killer feature is its gateway architecture that integrates with 50+ messaging platforms. The agent treats all connected channels as a unified conversation — a message from Telegram and one from Slack go to the same agent brain.

Telegram (Recommended for Beginners)

Telegram is the path of least resistance thanks to its official Bot API:

1. Chat with @BotFather on Telegram, run /newbot, and copy the token
2. Get your numeric Telegram ID from @userinfobot (send /start)
3. Add the token and your user ID to config.yaml
4. Send /start to your bot — Telegram bots can't initiate first contact
5. Start chatting; your agent responds autonomously

WhatsApp

WhatsApp has no bot token system. Instead, OpenClaw uses the Baileys library to connect via QR code:

1. Enable the WhatsApp channel in OpenClaw
2. Scan the terminal QR code from WhatsApp > Settings > Linked Devices
3. Message yourself — the agent responds to your own number

Security warning: OpenClaw recommends using a separate phone number for WhatsApp integration. Also, verify you're installing the genuine Baileys package — researchers discovered a malicious npm package called "lotusbail" that impersonated the WhatsApp library and stole credentials.

The Skill Ecosystem: ClawHub and AgentSkills

OpenClaw's extensibility comes from AgentSkills — a markdown-based plugin system that's refreshingly transparent. Each skill is a directory containing a SKILL.md file with YAML frontmatter (name, version, description, tags) and natural language instructions. No compiled code, no black boxes — you can read and modify every skill in a text editor.

ClawHub, the public registry, hosts over 13,729 skills with 1.5 million cumulative downloads. Installation is straightforward:

clawhub install 
clawhub update --all

Skills install to ./skills by default and are auto-detected on the next session. The AgentSkills spec is also supported by Claude Code, Cursor, and OpenAI Codex to varying degrees, though many ClawHub skills are written specifically for OpenClaw's gateway model.

Critical warning: A Bitdefender investigation found that 20% of ClawHub skills are malicious. Some were found silently exfiltrating data to external servers. Always review SKILL.md contents before installing, stick to high-download, positively-reviewed skills, and consider the VoltAgent/awesome-openclaw-skills curated collection on GitHub as a safer starting point.

How OpenClaw Compares: AutoGPT and LangChain

The three occupy different layers of the AI agent stack:

OpenClaw runs as a local gateway process. All orchestration — task planning, memory retrieval, tool invocation — happens on your hardware. It calls external APIs only for model inference. Instead of screenshot-based UI manipulation, it issues shell commands and API calls directly. This makes it faster and more token-efficient, but it also means the agent has direct access to your host system.

AutoGPT (181,000+ GitHub stars) runs in a Docker-sandboxed environment. Its isolation is a genuine security advantage — every reasoning step is contained. The trade-off is higher token consumption since every step requires an API call, and it lacks OpenClaw's native messaging integrations.

LangChain is a developer framework, not a finished product. It provides the building blocks for custom agent behavior — LLM calls, tool use, memory systems, multi-agent orchestration — but offers no out-of-box messaging channels, no skill marketplace, and no one-command install. It's the right choice when you need full architectural control.

For most users who want a practical, privacy-respecting AI assistant integrated into their daily messaging workflow, OpenClaw is the most accessible option today.

Security: The Elephant in the Room

OpenClaw's meteoric growth has outpaced its security maturity, and this cannot be sugarcoated.

CVE-2026-25253 (CVSS 8.8) is the headline vulnerability: a crafted malicious link can exfiltrate authentication tokens and give attackers full gateway control, including arbitrary command execution. Hunt.io found over 17,500 internet-exposed instances vulnerable to this exploit. Two additional command injection CVEs (CVE-2026-24763 and CVE-2026-25157) compound the risk.

Prompt injection remains an unsolved architectural challenge. Attackers can embed malicious instructions in web content that OpenClaw ingests, and the system currently cannot distinguish legitimate commands from adversarial prompts.

Shadow AI deployments are rampant: Token Security found that 22% of organizations have employees running OpenClaw without IT approval.

Microsoft's Security Blog, Cisco, Kaspersky, and multiple security firms have published advisories. The consensus recommendations:

• Keep OpenClaw updated — known CVEs are patched in recent releases
• Run under a dedicated non-privileged account with no access to sensitive data
• Set the allowedUsers field in config.yaml to restrict who can interact with the agent
• Review every ClawHub skill before installation
• In enterprise environments: deploy only in fully isolated containers or VMs with continuous monitoring
• Never expose your OpenClaw instance to the public internet without authentication

Getting the Most Out of OpenClaw: Practical Strategy

If you're starting fresh, here's a proven onboarding sequence:

Week 1: Install OpenClaw, connect Telegram (easiest channel), and use only built-in capabilities. Get comfortable with how the agent reasons, what it can and can't do, and how to phrase requests effectively.

Week 2: Add 3-5 highly-rated ClawHub skills from the curated awesome-openclaw-skills list. Start with productivity tools — calendar management, web search summarization, note-taking. Review each SKILL.md before installing.

Week 3: Connect a second channel (WhatsApp or Slack) and experiment with the unified conversation model. If API costs concern you, explore the Ollama integration for local model inference — performance won't match frontier models, but it's free to experiment.

Ongoing: Keep security hygiene as a habit. Update regularly, audit installed skills periodically, and follow the OpenClaw security advisories.

Looking Ahead

OpenClaw's trajectory — from a one-hour weekend prototype to GitHub's most-starred project to an OpenAI-backed foundation — encapsulates the AI industry's current velocity. The 250,000 stars, 48,274 forks, and 1,075 contributors represent something bigger than one project: they represent millions of people who are ready for AI that doesn't just talk, but acts.

The security challenges are real and serious. The malicious skill problem needs systemic solutions. But if the foundation governance holds, the security posture matures, and the community maintains its momentum, OpenClaw has a genuine shot at becoming the standard interface between humans and AI agents. The gap between "AI that chats" and "AI that works" is closing fast — and OpenClaw is leading the charge.