OpenClaw is the fastest-growing open-source project in history. 295,000 GitHub stars (basically bookmarks that developers use to mark projects as favorites) in less than 6 months. More than React (a widely used programming library by Meta) collected in 10 years.
I have been using OpenClaw since the early days, when it was still called Clawdbot. I went through the name changes, experienced the security vulnerabilities firsthand, and watched the insane hype unfold. From 60,000 stars in 72 hours to 1,000 people lining up outside Tencent headquarters in Shenzhen to get help installing the thing.
That said:
Not everything about OpenClaw is great. 36% of marketplace extensions contain prompt injections (an attack method where malicious instructions get smuggled into AI requests). Over 155,000 OpenClaw instances are sitting unprotected on the internet. And the founder left the project to work at OpenAI.
In this article, you will find all the latest numbers, data, and facts about OpenClaw. From GitHub stars and user numbers to security issues and enterprise adoption in China.
- 295,000+ GitHub stars (April 2026). OpenClaw remains the most starred software project on GitHub
- 38 million monthly website visitors, 3.2 million monthly active users, continued double-digit monthly growth
- Security concerns: 36% of ClawHub skills contain prompt injections, 155,000+ unprotected instances on the internet
1. What is OpenClaw?
OpenClaw is a free platform that connects various AI models (like ChatGPT, Claude, or Gemini) with messaging apps. You can use AI directly through WhatsApp, Telegram, or Slack. No separate app. No subscription. Just your existing messenger.
The idea behind it:
You install OpenClaw on a server (or locally on your machine), connect it to one or more AI models, and control it through your preferred messaging platform. The tool handles the orchestration. You just send a message.
The software is released under the MIT license. That means it is completely free, the source code is publicly available, and there are no restrictions on commercial use. You can copy OpenClaw, modify it, build it into your own product, and even sell it. No license fees. No permission needed.
The project launched in November 2025 under the name Clawdbot. Since then, it has had a meteoric rise:
In less than 6 months, it went from one Austrian developer's side project to the most starred software project on GitHub. Nothing like this has ever happened in the history of open source. And it shows just how massive the demand for a universal, open-source AI agent really is.
2. GitHub Stars
OpenClaw's growth curve on GitHub is unprecedented. No other software project in the platform's history has collected so many stars so quickly. The following chart shows the development since launch:
The key GitHub metrics at a glance:
For comparison:
React took over 10 years to reach 230,000 stars. OpenClaw did it in 60 days. On March 3, 2026, OpenClaw overtook React at exactly 250,829 stars to become the most starred software project on GitHub.
3. Website Visitors & Traffic
OpenClaw does not just attract developers. The website, OpenClaw.ai, now sees 38 million monthly visitors. The platform has 3.2 million monthly active users. And while growth is slowing down, it remains impressive.
3.1. Growth
In February/March 2026, traffic exploded by 925%. Nearly a 10x jump in a single month. In March/April 2026, traffic grew by another 41%. That sounds modest by comparison, but in absolute terms it is a leap from 27 to 38 million visitors.
For comparison:
ChatGPT hit 40% to 50% growth during its best months. OpenClaw is still running at the same level.
The following chart shows monthly visitor growth since launch:
3.2. User Behavior
45 minutes average session duration. That is unusually high. Not just for an open-source tool, but also compared to the major AI platforms:
gradually.aiOpenClaw users spend almost twice as much time on the platform as ChatGPT users. That tells you people are not just skimming the docs. They are actively working.
And the 92% retention rate shows:
Once someone sets up OpenClaw, they stick with it. That is an extremely high number for a tool that requires configuration and technical knowledge.
3.3. Desktop vs. Mobile
The OpenClaw website is visited predominantly on desktop. That makes sense because setup and administration happen primarily on a computer. The 20% to 30% mobile share likely comes from users accessing OpenClaw through WhatsApp or Telegram on their phones. That share will probably grow as more non-developers discover the tool.
gradually.ai3.4. Geographic Distribution
OpenClaw is used worldwide. The USA leads with 16.29% of traffic, closely followed by India and China. The fact that the top 3 countries are nearly tied is unusual. Most tech tools see the USA dominating by a much wider margin.
Particularly notable:
The explosive growth in China (+1,436%) and Canada (+1,259%). Both countries catapulted into the top 5 within a single month.
The growth rates are even more impressive. China and Canada broke into the top 5 within a single month:
China landing at #3 is remarkable. Despite the Great Firewall and its own AI ecosystems like Baidu and DeepSeek, OpenClaw has built a massive user base there. +1,436% growth in a single month. More on that in section 8.
Germany sits at #4 with 4.10%. At 38 million monthly visitors, that works out to roughly 1.6 million visits from Germany. With +992% growth, the German community will likely keep expanding quickly. The German-language documentation is still thin, though. Most guides (including my installation guide) are in German, but the official docs are English-only.
4. Which AI Models are Used with OpenClaw?
One of OpenClaw's biggest advantages:
It works with virtually every major language model and every messaging platform. That is fundamentally different from solutions like ChatGPT (OpenAI models only) or Claude (Anthropic models only). You choose which model you want. And you can switch anytime.
A particularly revealing look at actual usage comes from OpenRouter. OpenRouter is a routing platform for AI models that many OpenClaw users rely on to forward their requests to different AI providers.
4.1. Top 20 AI Models
The top 20 models by token usage show which AI models OpenClaw users rely on most:
4.2. AI Models by Provider
Four of the six largest providers are Chinese companies (stepfun, Xiaomi, z-ai/Zhipu, MiniMax). Together, they account for over 54% of total token usage.
Anthropic sits at #4 with Claude 4.6 Sonnet and Opus as the most-used Western models. Google follows at #6:
That lines up with the explosive China growth from section 5. Chinese users prefer local models that do not send data abroad. stepfun's Step 3.5 Flash alone consumes 3.474 trillion tokens, nearly matching Anthropic's entire model portfolio (2.135 trillion).
4.3. Token Usage of AI Models
Daily token usage in March 2026 shows the rapid rise of Chinese models:
In total, 19.2 trillion tokens (the smallest unit AI models use to process text, roughly comparable to word fragments) have been consumed by OpenClaw installations worldwide since January 2026. That is hard to wrap your head around.
For comparison:
That roughly equals 15 trillion words, or 150 times the entire text content of Wikipedia. In just 3 months.
4.4. Local AI Models
OpenClaw can also run entirely without cloud APIs. With Ollama, AI models run directly on your own machine. No data transfer to third parties. No monthly API costs. Full control over your data.
The most popular models for running OpenClaw locally (all values at Q4_K_M quantization, a 4-bit compression that reduces memory usage by roughly 75%):
Model | Parameters | RAM Required | Speed |
|---|---|---|---|
| Phi-4-mini | 3.8B | 4 GB | ~20 tokens/s |
| Mistral 7B | 7B | 5 GB | ~35 tokens/s |
| Llama 3.3 8B | 8B | 6 GB | ~30 tokens/s |
| Qwen 3 8B | 8B | 6 GB | ~28 tokens/s |
| Phi-4 14B | 14B | 10 GB | ~18 tokens/s |
| Qwen 2.5 Coder 14B | 14B | 10 GB | ~16 tokens/s |
| DeepSeek-R1 32B | 32B | 20 GB | ~10 tokens/s |
| Llama 3.3 70B | 70B | 42 GB | ~8 tokens/s |
| DeepSeek-V3 671B | 671B | 380 GB+ | ~2 tokens/s |
What hardware do you need for this? Macs are particularly well suited for local AI models because the GPU and CPU share the same memory (Unified Memory). That means the entire RAM is available for the model, not just the GPU memory.
Entry-level costs start at $800 for a Mac Mini M4 and go up to $8,000 for a research setup with 192 GB of RAM:
gradually.aiThe deciding factor is RAM. It determines how large a model you can run locally. From small 8-billion-parameter models all the way to frontier models with 405 billion parameters:
gradually.aiA breakthrough that could significantly lower hardware requirements soon:
Google introduced TurboQuant in March 2026 (to be presented at ICLR 2026). TurboQuant compresses the so-called KV cache (the working memory an AI model needs during a conversation) down to 3 to 4 bits. That reduces memory usage for long conversations by 4x to 6x.
In practice, that means:
Models that currently need 48 GB of RAM could run with 24 GB using TurboQuant. No fine-tuning required, no quality loss. Llama.cpp (the software foundation for Ollama) is already working on integration. The official Google implementation is expected in Q2 2026.
5. The OpenClaw Effect on Mac Sales
OpenClaw has not only changed the open-source world. It has changed the hardware market too. Because the tool runs best on Apple hardware (thanks to Unified Memory), demand for Mac Minis has surged since launch.
5.1. Delivery Times & Shortages
Delivery times on Apple.com show the scale of the demand:
In March 2026, the Mac Mini completely sold out across China. At Huaqiangbei market in Shenzhen (the world's largest electronics market), all Mac Mini stock was bought up. Retailers could not source replacements and had no idea when new shipments would arrive.
5.2. Price Increases & Resale Market
In Beijing and Shenzhen, retailers charged a premium of 500 yuan (roughly $73) on top of the list price. The resale market reacted too:
According to CNBC, used Mac prices on the ATRenew platform rose by around 15%. OpenClaw runs particularly well on Apple Silicon thanks to Unified Memory, and many users are buying Macs specifically to run the tool around the clock.
Apple CEO Tim Cook confirmed the trend during a China visit in March 2026. Apple integrated the Neural Engine into Macs over 10 years ago. With Apple Silicon, generative AI, and agent interactions, the Mac Mini is the best-suited computer for AI tasks.
5.3. Apple Quarterly Results
The numbers from Apple's latest quarterly report (Q1 FY2026, October to December 2025) show a strong Mac business:
The Q2 results (January to March 2026, the period when the OpenClaw hype peaked) are expected in late April 2026. Analysts anticipate a significant jump in Mac sales.
And: Apple announced in February 2026 that it will move Mac Mini production to the USA (Houston). Whether that is a response to the increased demand remains unclear. But the timing is conspicuous.
6. Community & Ecosystem
The OpenClaw community is not just growing in user numbers. The developer community behind it is equally impressive.
6.1. Contributors Compared
How quickly did OpenClaw build an active developer base compared to other major open-source projects? The difference is stark:
gradually.aiOpenClaw amassed nearly as many contributors in 6 months as React did in 10 years. Only the Linux Kernel and Kubernetes are clearly ahead, and they have existed for over a decade.
6.2. Weekly Development Activity
The GitHub activity shows how intensively the project is being worked on. In an average week during March 2026 alone:
gradually.aiDevelopment happens around the clock here. Literally, because contributors sit in every time zone.
6.3. Community Structure
The community organizes primarily through GitHub Issues, Discord, and a weekly open-source meeting. There are now also regional meetups in over 30 cities worldwide. In Berlin, Munich, and Vienna, monthly OpenClaw meetups take place. A sign that the German-speaking community is growing.
7. ClawHub Marketplace
ClawHub is the official marketplace for OpenClaw extensions (called "skills"). The numbers are impressive. But also concerning.
7.1. Growth
The marketplace grew from zero to over 44,000 extensions in just a few months. A growth rate that puts even established plugin ecosystems like the VS Code Marketplace to shame:
gradually.aiThe VS Code Marketplace took over 3 years to reach its first 10,000 extensions. ClawHub did it in 4 months.
7.2. Skill Categories
Which types of skills are developed most? The breakdown by category shows a clear focus on automation:
gradually.aiAutomation leads with over 10,000 skills. That lines up with the enterprise data from section 7, where 48% of users deploy OpenClaw primarily for productivity automation.
7.3. Developers & Installations
Behind those 44,000 skills sit 12,400 developers. Only 847 (6.8%) are verified by ClawHub. That means over 93% of all skill publishers have not provided proof of identity:
gradually.aiIn total, 2.3 million skill installations have been counted. The distribution is extremely uneven. The top 10 most popular skills account for 38% of all downloads:
gradually.ai7.4. Quality Issues
Here is the thing:
The ecosystem is growing faster than the security measures can keep up. That is typical for open-source projects in hypergrowth mode. But for a tool that has access to AI models and messaging platforms, it is a serious problem.
The fact that skill counts vary between 20,000 and 44,000 depending on the source reveals another problem:
There is no official, reliable count. Some skills are listed twice. Others are forks of existing skills with minimal changes. The actual number of unique, working skills is likely much lower.
8. Security & Vulnerabilities
This is the part nobody likes hearing. But it matters. And it gets overlooked in the euphoria around GitHub stars.
OpenClaw has serious security problems. Not because the developers are doing sloppy work. Because the project grew so fast that security measures simply cannot keep up. 145,000 lines of code, over 1,200 contributors, thousands of marketplace extensions. Building all of that in less than 6 months while keeping it secure is practically impossible.
8.1. Known Vulnerabilities
Security Issue | Date | Status |
|---|---|---|
| Fehlende Security-Reviews | 22. Mär. 2026 | In Arbeit |
| Exponierte Instanzen | 21. Mär. 2026 | Nutzerproblem |
| ClawHub Prompt Injections | 17. Mär. 2026 | Ungelöst |
| Token-Leaking über Logs | 15. Mär. 2026 | Teilweise behoben |
| CVE-2026-31205 | 9. Mär. 2026 | Gepatcht |
| Fehlende Rate-Limiting | 8. Mär. 2026 | In Arbeit |
| Unverschlüsselte API-Keys | 3. Mär. 2026 | Gepatcht (v0.8.2+) |
| CVE-2026-27841 | 28. Feb. 2026 | Gepatcht |
| Unsichere Default-Config | 14. Feb. 2026 | Doku aktualisiert |
| CVE-2026-25253 | 12. Feb. 2026 | Gepatcht |
Security researchers from Cisco, Microsoft, and Kaspersky have independently flagged these issues. The OpenClaw Foundation is working on a security review process for the ClawHub Marketplace. But so far, there is no mandatory review.
8.2. ClawHub Marketplace Security
The security situation in the marketplace is alarming. An analysis by security researchers shows the breakdown of problems:
gradually.aiOnly 47% of skills are safe. 36% contain prompt injections (hidden instructions that manipulate the AI). 8% actively attempt to send user data to external servers. 6% request permissions far beyond their actual functionality.
The critical vulnerability CVE-2026-25253 was particularly alarming. An attacker could execute arbitrary code on the server through a manipulated message. A single click was enough. The vulnerability was patched within 48 hours. But how many of the 155,000+ unprotected installations were compromised in the meantime remains unclear.
155,000 unprotected installations. Let that sink in. That means roughly one in every fifteen OpenClaw instances is sitting on the internet without basic access protection. Many users install the tool, set it up, and forget to configure the firewall.
My advice:
Do not use OpenClaw with sensitive data. Keep it up to date at all times. And do not install skills whose source code you have not reviewed.
If security is your top priority, check out the most secure OpenClaw alternatives. OpenFang and IronClaw are built specifically for security-critical environments and have a significantly smaller attack surface.
9. The China Wave
Particularly remarkable:
The adoption in China. Tencent, Alibaba, ByteDance, Baidu, and Xiaomi are all using OpenClaw. These are not small startups. They are the five largest technology companies in China.
One image went viral in early March 2026:
Over 1,000 people outside Tencent headquarters in Shenzhen, waiting for help installing OpenClaw. In a line. For an open-source tool.
That tells you two things:
How massive the demand is. And that the installation process is still too complicated for many users.
The growth in China (+1,436% MoM) explains a large chunk of the global traffic. China alone accounts for 12.08% of all OpenClaw visits. And that is despite the fact that many of the supported AI models (like Claude or GPT) are not directly available in China.
Chinese companies primarily use OpenClaw with local models via Ollama, DeepSeek, or Chinese cloud providers like stepfun, Zhipu AI, and Xiaomi. The OpenRouter statistics confirm that.
Four of the six most-used providers are Chinese. stepfun's Step 3.5 Flash alone consumes more tokens than all Anthropic models combined. Combined with OpenClaw, that creates a completely local, privacy-compliant AI agent. No data leaving the country, no third-party subscription.
For Western observers, this is remarkable. While Europe is still debating AI regulation, Chinese mega-corporations have already integrated OpenClaw into their internal workflows. The speed of adoption in Asia shows just how large the global demand for open AI agent frameworks really is.
10. Founder & History
The person behind OpenClaw is Peter Steinberger from Austria. An experienced iOS developer who previously founded and led PSPDFKit. PSPDFKit was one of the most successful PDF frameworks in the Apple ecosystem.
Steinberger started OpenClaw as a side project. Evenings and weekends, alongside his main job. Six months later, it is the most starred software project on GitHub. That is the kind of story that only happens in the open-source world.
The founder joining OpenAI while his open-source project is breaking every record? That triggered mixed reactions in the community.
That said:
The OpenClaw Foundation ensures the project lives on independently. The MIT license guarantees that nobody (including OpenAI) can restrict its use. And with over 1,200 active contributors, the project is no longer dependent on any single person.
The trademark issues are an interesting story too. Anthropic (the company behind Claude) filed a complaint about the name "Clawdbot" because of the phonetic similarity to "Claude." The rename to "Moltbot" lasted only a few weeks. Then came "OpenClaw" with a foundation structure that put the project on a more stable legal footing.
3 names in 4 months. Normally, that would be a branding disaster. But the community followed every name change without user numbers dropping. That shows how strong the attachment to the product is, not the brand.
Steinberger's move to OpenAI also raises the question of how much influence OpenAI will have on OpenClaw's development going forward. The Foundation emphasizes there is no connection. But when the creator of an open-source project works at one of the largest AI companies, skepticism is warranted.
11. Messaging Platforms
OpenClaw officially supports 8 messaging platforms. On top of that, there are 9 community integrations developed and maintained by the open-source community.
Platform | Status | Images | Voice | Groups | E2E | Setup |
|---|---|---|---|---|---|---|
| Official | ✓ | ✓ | ✓ | ✓ | Medium | |
| Telegram | Official | ✓ | ✓ | ✓ | ✗ | Easy |
| Slack | Official | ✓ | ✗ | ✓ | ✗ | Medium |
| Discord | Official | ✓ | ✓ | ✓ | ✗ | Easy |
| Signal | Official | ✓ | ✓ | ✓ | ✓ | Hard |
| iMessage | Official | ✓ | ✗ | ✓ | ✓ | Hard |
| Google Chat | Official | ✓ | ✗ | ✓ | ✗ | Medium |
| MS Teams | Official | ✓ | ✗ | ✓ | ✗ | Hard |
| Messenger | Community | ✓ | ✗ | ✓ | ✗ | Medium |
| Community | ✓ | ✓ | ✓ | ✗ | Hard | |
| Matrix | Community | ✓ | ✗ | ✓ | ✓ | Medium |
| Mattermost | Community | ✓ | ✗ | ✓ | ✗ | Easy |
| Rocket.Chat | Community | ✓ | ✗ | ✓ | ✗ | Easy |
| Community | ✓ | ✗ | ✗ | ✗ | Easy | |
| SMS | Community | ✗ | ✗ | ✗ | ✗ | Medium |
| IRC | Community | ✗ | ✗ | ✓ | ✗ | Medium |
| Threema | Community | ✓ | ✓ | ✓ | ✓ | Medium |
The estimated setup time varies significantly by platform:
gradually.aiThe variety of integrations is one of the main reasons for its success. You do not need to switch between platforms. You just use the messenger you already have. WhatsApp alone has over 2 billion users worldwide. That lowers the barrier to entry enormously.
Compared to Claude Code (which is a pure command-line tool), OpenClaw is far more accessible for non-developers. You do not need a terminal. You just send a message. Your grandmother could theoretically use Claude through WhatsApp without knowing what a terminal is. That is the strength of the messenger approach.
For a detailed setup guide, check out my OpenClaw installation guide.
12. Comparison with Alternatives
OpenClaw is not the only open-source framework for AI agents. There is a growing landscape of alternatives with different focus areas. Some prioritize security. Others speed. Others a minimal codebase.
The comparison by GitHub stars shows how far ahead OpenClaw is. But also that the market for AI agent frameworks is exploding across the board:
OpenClaw dominates in popularity. But popularity alone says nothing about quality or security. OpenFang, for example, is built specifically for security-critical environments and is regularly audited by external security experts. Hermes encrypts all communication between components automatically. And Nanobot gets by with just 3,000 lines of code instead of OpenClaw's 145,000.
For regulated industries (healthcare, finance), IronClaw with its government-recognized security certification is the better choice. For teams that want to start fast, ZeroClaw is ready in under 2 seconds. And if you need maximum performance with minimal resource usage, Nanobot is the way to go.
The market for AI agent frameworks is still very young. In 12 months, the landscape will probably look completely different. New frameworks will appear, existing ones will merge or disappear. But as of today, OpenClaw leads in popularity while the alternatives have the edge in specific areas (security, performance, simplicity).
For a detailed comparison, check out my article on the best OpenClaw alternatives.
13. Conclusion
OpenClaw is an impressive project with an unprecedented growth story. 295,000 stars. 38 million monthly visitors. 19.2 trillion tokens on OpenRouter. Adoption by the largest tech companies in the world.
But it is not a perfect tool. The security issues are real. The ClawHub Marketplace urgently needs better quality controls. And the question of how the project will evolve after the founder's departure remains open.
My assessment:
OpenClaw is here to stay. The community is too large, the usage too broad, the ecosystem too vibrant for the project to simply disappear.
But the next 6 months will be decisive. If the Foundation gets the security issues under control and the ClawHub Marketplace introduces mandatory security reviews, OpenClaw has the potential to become the standard framework for AI agents. If not, the security-focused alternatives like OpenFang and Hermes will take over.
I will keep updating this article as new numbers become available. Development at OpenClaw moves so fast that today's statistics could be outdated within a month.
If you want to get started, check out my installation guide. If you want to compare alternatives first, I have put together the 10 best OpenClaw alternatives. And for a general overview of AI chatbots and the best AI tools, take a look at my other articles.
