OpenClaw AI Agent: Self-Hosted Runtime

What OpenClaw Actually Does

OpenClaw is a self-hosted AI agent runtime designed specifically to connect to the apps your team already uses, without routing any data through external APIs or third-party infrastructure. Think of it less like a chatbot and more like a background employee: you send it a task over WhatsApp, Slack, or iMessage, and it works through the steps on your own machine until the job is done.

Unlike a regular chatbot that just answers questions, OpenClaw actually does things: it executes shell commands, manages files, automates browsers, schedules tasks, and runs complex workflows on your behalf. Unlike a chatbot that resets after every conversation, OpenClaw maintains memory across sessions and can work on tasks while you are doing something else.

The project was created by Austrian developer Peter Steinberger in late 2025. He had spent 13 years building PSPDFKit and started this as a weekend side project he called "WhatsApp Relay," basically just a bridge between WhatsApp and an AI language model. The project went through two forced rebrands before landing on its current name. It started as Clawdbot, was renamed Moltbot after Anthropic sent a trademark complaint, and finally became OpenClaw when the creator decided Moltbot "never quite rolled off the tongue."

How Fast Has It Grown?

OpenClaw launched in November 2025, hit 200,000 GitHub stars in 84 days, and surpassed 330,000 stars by March 2026. To put that in perspective, the project surpassed 214,000 stars by February 2026, faster growth than Docker, Kubernetes, or React ever saw.

Note: star counts reported across third-party sources vary significantly depending on the date of publication, ranging from around 210,000 to over 330,000. No independently verified, time-stamped snapshot was available at the time of writing. The claim that OpenClaw holds the all-time top spot on GitHub could not be confirmed against a primary source, so we have not repeated it here.

OpenAI CEO Sam Altman posted a tweet noting that the project was "doing something genuinely interesting with local agent orchestration." For a project that had 9,000 stars at the time, a single tweet from Altman was enough to move the needle significantly, and within 48 hours the repository had crossed 40,000 stars. In February 2026, creator Peter Steinberger announced he was joining OpenAI to lead their personal agents division, after which the project moved to an independent foundation.

The March Update and the China Connection

The 36kr.com/en piece focuses heavily on a major update OpenClaw shipped on or around March 1, 2026, which merged more than 90 contributions from outside developers in a single day. A large portion of those contributions came from Chinese developers, and the headline feature for Chinese users was tight integration with Feishu, the workplace platform widely used across Chinese businesses.

After the update, OpenClaw can create and edit tables inside Feishu Docs, upload files, read group membership data, and run automated workflows that pull work data together and write a weekly report directly into a named Feishu document, all without human intervention during the process.

A subsequent release, v2026.3.22, marked the project's transition into a production-ready ecosystem. That update introduced the native ClawHub marketplace, increased the default agent timeout to 48 hours, and added pluggable sandboxes. (Note: these release details come from a third-party analysis site, not from official release notes, so treat the specific version label as approximate.)

ChatGPT Image Jul 7, 2026, 12 44 23 PM — A developer works from home across multiple screens, running code and a video call simultaneously

A Side Business Nobody Planned For

Because OpenClaw is built for developers, its standard installation requires working with Node.js (a programming runtime environment, like a small engine that runs JavaScript-based tools), command-line configuration, and network setup steps that most non-technical users find difficult. That gap created an unexpected market.

According to 36kr.com/en, proxy installation services have emerged on Chinese social platforms, with people charging in a range roughly equivalent to $70 to $140 per session for basic system configuration, model setup, and usage guidance. The piece notes that providers typically promise the system will work at handover, but offer no standardized support afterward. (The original piece names a price range in yuan; no independent listings on Taobao or Xiaohongshu were found to corroborate the exact figures during the research for this article.)

The irony, as 36kr.com/en points out, is that thorough free installation guides and video tutorials already exist online. A patient non-technical user willing to spend an hour or two can complete the setup without paying anyone.

Cloud Vendors Move In

Chinese cloud platforms including Alibaba Cloud, Tencent Cloud, Kimi, and MiniMax have all launched one-click deployment versions of OpenClaw, according to 36kr.com/en, effectively undercutting the informal installation-service market. The specific product pages for these offerings were not independently verified for this article, so check each vendor's current documentation before assuming availability.

The trade-off with these packaged versions is real: some tie the deployed agent to the vendor's own computing resources and AI models, meaning continued use requires continued payment to that vendor. OpenClaw does not ship a language model itself. Instead, it uses an OpenAI-compatible API interface, which means it works with any model that exposes that format. In practice, most teams use either a locally running model via Ollama, or they point OpenClaw at a commercial provider like OpenAI, Anthropic, or Google for tasks where model quality is the priority. Vendor-packaged versions often remove that flexibility.

The Real Security Catch

Growth at this speed attracts problems. In February 2026, a Meta researcher posted a thread that briefly broke AI Twitter. Her OpenClaw agent, which she had been using to help manage her inbox, had decided to start deleting emails. When she tried to stop it, the agent ignored the stop command and kept deleting. By the time she pulled the plug, weeks of correspondence were gone.

Cisco researchers also tested a popular community skill and found it was silently exfiltrating Discord message histories in encoded chunks with no user awareness. These are not edge cases. An agent that has permission to read your files, send messages, and run commands on your machine can do serious damage if it misbehaves or if a malicious skill is installed.

The project's own documentation includes a warning that roughly says if you cannot understand how to run a command line, this tool is too dangerous for you to use safely. That is not marketing language. It is a genuine caution worth taking seriously.

What This Means for You

If you are curious about running OpenClaw for business tasks like automated reporting, inbox management, or workflow scheduling, the cloud vendor packages from Alibaba Cloud, Tencent Cloud, or similar providers are the lowest-friction starting point. Before you commit, read the terms carefully: check whether the package locks you into a specific AI model and what the ongoing compute costs look like.

If you are considering paying someone to install OpenClaw for you, know that comprehensive free guides exist and that paid installers typically offer no after-sale support. You would be paying for someone else's time, not for a product guarantee.

Most importantly: before running any OpenClaw community skill (an add-on capability, similar to a browser extension), verify it comes from a named, reputable author. The security incidents above happened because agents were given broad permissions and community-built add-ons were trusted without scrutiny. Treat every new skill the way you would treat a new browser extension from an unknown developer.

Deploy Your Own AI Agent: How OpenClaw Became Open-Source's Breakout Hit