Meta’s AI safety executive watched helplessly as AI deleted all their emails. The cyber bandit responsible was an open-source project only 60 days old – OpenClaw. In just 60 days, OpenClaw’s GitHub popularity surpassed Linux and TSFlow, projects with over a decade of legacy. Some see it as an automated money-printing machine, earning $110,000 weekly. Yet some users accidentally exposed their account credentials and full chat records.
Ladies and gentlemen, please take your seats. Fellow enthusiasts, please take your seats. OpenClaw’s explosive popularity in early 2024 has sparked debate: is it AI agents’ future or a unprecedented trap? Remarkably, OpenClaw began as a one-hour weekend project by Austrian programmer Peter. He merely wanted to connect his 24/7 working CloudCode AI to messaging apps, so he could control it remotely. But this lighthearted project soon gave him cold sweats.
One day, Peter sent an audio message to AI, realizing too late he hadn’t enabled voice recognition. Suddenly his phone buzzed – AI responded! Checking the backend, he found AI had self-written voice-to-text functionality. This is the essence of AI agents. While agent capabilities aren’t new, Peter’s OpenClaw brought them into mainstream view. Unlike complex command-line tools like CloudCode (used by programmers), OpenClaw bridges these agents with WhatsApp, iMessage, Discord (20+ platforms), making AI controllable via mobile messages on Macs and Androids.
In practice tests, I first asked OpenClaw to clean my browser history. During family gatherings, relatives’ kids often accessed adult content. OpenClaw didn’t just delete history – it directly removed Chrome’s local history files, cookies, and cache in one go. I also tested file organization and email sorting. OpenClaw categorized 100+ emails with directory suggestions. For WeChat work tasks, you just issue commands and wait for results – no need to monitor processes.
This paradigm shift contrasts traditional software logic. Previously, we competed on service quality and interface design. Now AI agents skip the “plowing” step – you just say “plow the field” and wait for harvest. OpenClaw’s creator claims 80% of apps will either die or become AI agent interfaces. But risks emerge: when you issue a command in chat, OpenClaw executes autonomously. You don’t know if it’s plowing or detonating a nuke in your field.
I tested OpenClaw’s machine learning course download. It first tried keyword search (blocked), then used ytdlp (failed). After self-correcting and finding existing downloads, it finally succeeded. The backend showed it cycles through: 1) tool execution, 2) result verification. Tool selection and judgment accuracy depend on model strength – we chose MiniMax M2.5 balancing performance and cost.
OpenClaw’s core tool is axy (executing terminal commands), akin to CloudCode’s spirit. Other tools: read (file access), process (task tracking), sesence history (drawing history). Above these sits the skills ecosystem – pre-written prompt sets teaching AI specific tasks. For weather skills, it details usage scenarios and location queries. During tasks, agents check available skills and follow instructions step-by-step.
This skills ecosystem enables global AI enthusiasts to share network commands and prompt techniques as sharable documents. Even agents can create skills – like when OpenClaw wrote its own browser history cleanup skill. GitHub’s “awesome open class skills” project lists 2,800+ skills across finance, marketing, gaming. The clouhub platform hosts nearly 6,000 skills.
In The Matrix, data injection taught Neo martial arts. Today, agents achieve this, but dangers lurk. After integrating OpenClaw with Feishu, colleagues extracted API keys, gateways, and credentials. A full-access agent could easily leak sensitive info.
Current agents struggle with context limits. For complex tasks with many steps, agents often forget earlier instructions. The Meta executive’s email deletion incident also stemmed from forgotten initial commands. Solutions involve information curation – selecting partial tool lists, compressing memories. This causes unstable performance: same webpage analysis task yields different results based on provided tools.
Until context memory improves, agents remain “step-by-step newbies.” Asking for奶茶 (milk tea) might lead to: 1) hailing a ride, 2) installing payment apps, 3) awkwardly returning with milk tea. Thus, OpenClaw suits simple repetitive tasks best – a 24/7 diligent AI butler.
OpenClaw’s success lies in its perfect timing and form. CloudCode and skills concepts existed before, but AI’s current momentum made OpenClaw inevitable. It demonstrates the sci-fi dream: sending a mobile message to automate any task worldwide. As Peter says, OpenClaw isn’t magic – each step is mundane. But sometimes, recombining existing elements with new ideas creates magic. Regardless, OpenClaw marks a milestone. In this geek era, technology isn’t just big companies’ privilege. With OpenClaw, everyone can have their own Jarvis. While many challenges remain, welcome to the age of intelligent systems.