What is OpenClaw?
OpenClaw is a free, open-source autonomous AI agent. Created by Austrian developer Peter Steinberger and released under the MIT licence, it connects large language models to your real-world systems and executes tasks on your behalf through messaging platforms you already use: WhatsApp, Telegram, Slack, Microsoft Teams, Discord, Signal, iMessage, and more than a dozen others.
The critical distinction between OpenClaw and an AI chatbot is that OpenClaw does not just answer questions. It acts. It can read your inbox and summarise it, book calendar events, control a browser, execute shell commands, write files to disk, trigger workflows, submit forms, and complete multi-step processes from a single instruction. You message it the way you would message a colleague, and it works.
The project was originally published as Clawdbot in November 2025, briefly renamed Moltbot in January 2026 following a trademark dispute, and settled on OpenClaw three days later. By March 2026 it had become a cultural phenomenon in China, earned a Jensen Huang endorsement, and triggered a gold rush of installation service businesses across Asia.
"It is now the largest, most popular, the most successful open-sourced project in the history of humanity. This is definitely the next ChatGPT."Jensen Huang, CEO NVIDIA, March 2026
How the OpenClaw gateway works
The OpenClaw gateway is the central process that makes the system function. When you run OpenClaw on a machine, the gateway starts as a local WebSocket server on port 18789. Everything passes through it: incoming messages from your connected platforms, instructions to the agent runtime, responses from the LLM, and skill execution calls.
By default the gateway binds to the loopback address (127.0.0.1), meaning it only accepts connections from the local machine. This is a deliberate security boundary. Remote access is handled via SSH tunnelling rather than exposing the gateway directly to the network.
The architecture has three main layers working together:
- Gateway: owns all channels, authenticates connections, routes messages, and dispatches tasks to the agent runtime.
- Agent runtime: interprets instructions, loads your workspace context (defined in a SOUL.md configuration file), and invokes the appropriate tools or skills to complete the task.
- LLM provider: the reasoning engine. This can be a local Ollama model (recommended for privacy), the Anthropic API, OpenAI, DeepSeek, or any OpenAI-compatible endpoint.
When you send a message via Telegram asking OpenClaw to draft a reply to an email and add the sender to your CRM, the gateway receives it, the agent runtime interprets the instruction using the LLM, and the skills for Gmail and your CRM execute the actual actions. The result comes back to you in the same chat.
OpenClaw with Ollama: fully local, zero API cost
One of the most important configurations for business use is connecting OpenClaw to a local Ollama instance. Ollama runs open-weight language models on your own hardware, exposing an OpenAI-compatible API at port 11434. When configured, OpenClaw auto-discovers all available models and routes every reasoning task locally. No data leaves your network. No API invoice arrives at the end of the month.
For this to work well, the model needs a minimum 64k token context window to handle multi-step tasks reliably. A 20-billion-parameter model running on appropriate hardware (such as the DataForge AI Server with 16 GB VRAM) handles this comfortably at around 31 tokens per second, which feels real-time for most business workflows.
Running OpenClaw on your own hardware?
The DataForge AI Server comes pre-loaded with Ollama and the right models. The ForgeBox runs OpenClaw as its agent. Both are delivered, configured, and integrated by us.
How to install OpenClaw
The OpenClaw install process is straightforward for developers. The full setup takes around 10 minutes on a supported system.
Install Node.js 22 or later on macOS, Linux, or Windows via WSL2. Native Windows is not supported.
Install the OpenClaw CLI globally via npm. This downloads the gateway and agent runtime.
Run the onboarding command. OpenClaw walks you through a setup sequence in the terminal.
Select your LLM provider. Choose Ollama for local inference, or connect an API key for a cloud provider.
Configure your messaging channels. Connect Telegram, WhatsApp, Slack, Teams, or whichever platform your team uses.
The gateway starts automatically at 127.0.0.1:18789. OpenClaw is live and ready to receive instructions.
Hardware requirements are modest. OpenClaw itself uses 200 to 400 MB of RAM at idle. It runs on a modern laptop, a NUC mini PC, or a Raspberry Pi-class device. If you are running Ollama alongside it for local LLM inference, you need significantly more: a dedicated GPU with at least 8 GB of VRAM is recommended for business-quality model performance.
OpenClaw on GitHub: the numbers behind the hype
The OpenClaw GitHub repository sits at github.com/openclaw/openclaw. As of mid-March 2026 it had accumulated over 61,000 stars, 47,700 forks, and contributions from more than 600 developers. The codebase is TypeScript, MIT-licenced, and actively maintained.
To put the growth in context: React, one of the most widely used JavaScript frameworks ever created, took 10 years to reach a comparable star count. OpenClaw covered the same distance in 60 days. On March 3, 2026, it surpassed React's all-time record and became the second most-starred non-aggregator software project on GitHub, behind only TensorFlow.
The related repositories include the skills registry (openclaw/skills), a Nix package, and a Windows companion suite with system tray integration and PowerToys extension.
What the OpenClaw app does for your business
The OpenClaw app layer, meaning the skills and workflows you build on top of the agent, is where the business value is realised. The most common deployments NZ and AU businesses are exploring fall into five categories.
Inbox and communications
OpenClaw reads your inbox each morning, categorises by priority, drafts replies for your review, and logs all actions in your CRM without manual input.
Scheduling and calendar
Time-blocks tasks by priority, manages scheduling conflicts, books meetings, and sends daily briefings with objectives and key events.
Client onboarding
On contract sign, triggers a full onboarding sequence: document requests, follow-up reminders, internal Slack notifications, and checklist tracking.
Development and infrastructure
Triggers deployments from chat, monitors logs, identifies error root causes, writes and commits code fixes, and runs the full dev lifecycle autonomously.
Sales and reporting
Monitors pipeline, pulls CRM data, generates weekly summaries, tracks competitor content, and surfaces opportunities without opening a dashboard.
Customer support
Monitors support channels, triages tickets by urgency, drafts responses from your knowledge base, and escalates with full context when human input is needed.
Why OpenClaw news is everywhere right now
Three events converged in March 2026 to push OpenClaw from developer tool to mainstream business conversation.
China's "raise a lobster" movement. Chinese tech companies began deploying OpenClaw at scale in early March. Tencent launched an internal "lobster special forces" suite built on the framework. JD.com began offering remote installation services for 399 yuan. Engineers were charging 500 yuan for on-site setups. On March 11, nearly a thousand people queued outside Baidu headquarters in Beijing to get it installed. The cultural moment generated global media coverage that introduced OpenClaw to millions of business owners who had never heard of it.
Meta's acquisition of Moltbook. On March 10, Meta acquired Moltbook, the AI-agent social network that emerged from OpenClaw's community. The acquisition signalled that the largest technology companies view the agentic communication space as a serious long-term platform, not an experiment.
Jensen Huang's public endorsement. On March 17, NVIDIA's CEO called OpenClaw the most successful open-source project in history and the next ChatGPT. NVIDIA simultaneously announced NemoClaw, an enterprise security wrapper around OpenClaw for organisations that need governance controls before deployment.
The security risk you need to know before you install OpenClaw
OpenClaw's growth has outpaced its security maturity. Before deploying it in any business context, these risks need to be understood and mitigated.
CVE-2026-25253 (CVSS 8.8, Critical): A one-click remote code execution vulnerability was disclosed in February 2026. The gateway's Control UI trusts a gateway URL parameter without validation, allowing an attacker to send a crafted link that causes the victim's browser to automatically transmit the gateway authentication token. With that token, an attacker has full system access. Over 30,000 exposed instances were identified. All versions before 2026.1.29 are affected. Update immediately if you are running an older version.
A further concern is the skill ecosystem. Security researchers found that approximately 20% of published skills in the ClawHub registry contained malicious code. Installing skills from unknown publishers carries real supply-chain risk. For business deployments, only install skills from verified sources or build your own.
This is one of the primary reasons DataForge wraps OpenClaw deployments in a governance layer. We vet skills, harden the gateway configuration, manage updates, and ensure the agent only has access to the systems its role requires.
Want OpenClaw set up securely for your business?
DataForge configures, hardens, and integrates OpenClaw as part of the ForgeBox deployment. We handle the security so you can focus on the outcomes.
How OpenClaw compares to other AI agent frameworks
| Framework | Best for | Interface | Setup complexity |
|---|---|---|---|
| OpenClaw | Personal and team automation via messaging | WhatsApp, Telegram, Slack, Teams | Low (config-driven) |
| CrewAI | Multi-agent enterprise workflows | API / code | Medium (Python) |
| LangGraph | Stateful, predictable enterprise flows | API / code | High (flow design) |
| AutoGPT | Autonomous goal-breaking and exploration | Web UI / API | Medium |
| Claude Code | Development, infrastructure, coding tasks | Terminal / Slack bridge | Low (CLI-based) |
OpenClaw and Claude Code are complementary rather than competing. OpenClaw is the conversational interface that staff use through familiar messaging apps. Claude Code, running on a ForgeBox, is the development and infrastructure agent that executes technical work. Both can operate in the same organisation, serving different functions and talking to the same AI Server.
What this means for New Zealand businesses
The Google Trends data tells the story clearly. Every query related to OpenClaw, from "what is openclaw ai" through to "openclaw gateway" and "install openclaw", is recording breakout status in March 2026. That means search interest has grown more than 5,000% compared to baseline. New Zealand business owners and IT decision-makers are searching for this information right now.
The businesses that act first will gain the operational advantage that comes from deploying agentic AI while competitors are still researching it. The businesses that move without governance will encounter the security risks that are already documented. The middle path is a structured deployment with someone who has already solved the security, integration, and governance problems.
That is what DataForge provides. We deploy OpenClaw and Claude Code on ForgeBox hardware, connect it to your business systems, harden the security configuration, and train the agent on your specific context. You get the capability without the risk of a DIY installation.