OpenClaw: How an AI Agent Can Work Across Your Computer

OpenClaw is an open-source AI agent designed to do more than answer prompts: with permission, it can work with files, messaging clients and tools on a user's own computer. In a February 2026 interview with Lex Fridman, creator Peter Steinberger discusses how the project evolved from earlier names including MoltBot and Clawdbot and rapidly attracted attention on GitHub.

Its appeal is the idea of a model-agnostic assistant that can take actions across a local computing environment rather than remaining inside a chat window.

The core innovation of OpenClaw lies in its ability to bridge the gap between language and agency, transforming ideas into actions. It's an AI that 'gets you' and learns from you, all within an open-source, community-driven framework. This powerful capability, however, comes with significant responsibility. By granting system-level access, users gain immense power but also face potential security minefields.

OpenClaw represents a future where AI can be an extremely useful personal assistant, but this freedom demands careful consideration of data protection and cybersecurity threats.

Peter Steinberger, the creator of OpenClaw, is no stranger to building impactful software. He previously spent 13 years developing PSPDFKit, a tool used on over a billion devices. After selling it, he took a three-year hiatus before rediscovering his passion for programming and rapidly building OpenClaw. His journey embodies the spirit of the AI revolution in the programming world, marking the 'OpenClaw moment' as a significant milestone alongside the ChatGPT and DeepSeek moments.

The genesis of OpenClaw can be traced back to Steinberger's desire for a personal AI assistant. He experimented with pulling in data from WhatsApp and running queries on it, achieving profound results that even brought friends to tears. However, he initially believed that major labs would develop such tools. It wasn't until November that, frustrated by the lack of such an assistant, he decided to 'prompt it into existence.'

This mirrors his earlier entrepreneurial spirit with PSPDFKit, where a personal need (the inability to easily display PDFs on an iPad) led to a billion-device product.

The initial prototype, built in just one hour, involved hooking up WhatsApp to a cloud code environment. This allowed for interaction via CLI messages, with the agent performing tasks and returning results. While rudimentary, it felt 'really cool', the ability to 'talk to my computer.' The addition of image support, which took a few more hours, significantly enhanced its utility, allowing users to screenshot posters or events and get immediate analysis.

This early version, tested extensively during a trip to Marrakesh, proved its value by functioning even with shaky internet, translating, explaining, and finding places, essentially acting as a personal Google.

The experience of interacting with an AI through a chat client, rather than a traditional terminal, represents a 'phase shift' in how AI integrates into daily life. Steinberger likens the magic of this experience to the fundamental pleasantness of scrolling on an iPhone, components that were there, but brought together in a revolutionary way.

A pivotal moment for him was when a typing indicator appeared after he sent an audio message, an unexpected capability that the agent figured out on its own. The agent identified the audio file format, used ffmpeg for conversion, and leveraged OpenAI's Whisper and Curl to translate it, demonstrating remarkable problem-solving and world knowledge.

The journey of OpenClaw has not been without its challenges, particularly concerning its name. Initially named Wa-Relay, then Claude's, and later Clawdbot, the project faced a significant hurdle when Anthropic, the creators of the AI model Claude, requested a name change due to potential confusion. This led to a frantic renaming process, complicated by the explosive growth of the project and attempts by crypto enthusiasts to 'snipe' domain names and social media handles.

The saga involved multiple name changes, including a brief stint as Mod Bot, and significant effort to ensure an atomic, system-wide update across all platforms.

The name change process was fraught with difficulties, including domain squatting and malware distribution attempts by malicious actors. Steinberger recounts the stressful experience of trying to rename accounts and packages atomically, only to have names 'sniped' within seconds. This ordeal highlighted the aggressive tactics of certain online communities and the technical complexities of managing a rapidly scaling open-source project.

Despite these setbacks, the project eventually settled on OpenClaw, a name that, while perhaps not perfect, allowed the project to move forward.

Beyond the naming saga, OpenClaw's architecture involves several key components: a gateway, chat clients (like Telegram, WhatsApp, Signal), a harness, and the agentic loop itself. Steinberger emphasizes that building an agent loop is a fundamental 'Hello World' for AI, surprisingly simple and achievable. He also introduced a 'heartbeat' mechanism, initially a simple 'surprise me' prompt, which later evolved to be more context-aware, even checking in on him after surgery.

While critics might dismiss this as a mere cron job, Steinberger sees it as a way to make the AI more relatable and proactive, demonstrating a deeper level of interaction beyond simple task execution.

The rise of AI agents like OpenClaw prompts a fundamental question: what is the future of programming? Steinberger suggests that while AI may eventually replace some aspects of traditional programming, the art of building and conceptualizing (understanding what to build, how it should feel, and the overall architecture) will remain crucial. He likens the future of pure coding to a craft like knitting, pursued for enjoyment rather than necessity.

This shift, while potentially painful for those deeply identified as programmers, opens new avenues for creativity and problem-solving.

Steinberger advocates for 'agentic engineering,' a paradigm where humans collaborate with AI agents. This involves learning to communicate effectively with agents, understanding their strengths and limitations, and guiding them through complex tasks. He emphasizes that while AI can handle much of the intricate coding, the human role shifts towards higher-level design, problem-solving, and ensuring the 'delight' and 'soul' of the software.

This collaborative approach allows builders to focus on more impactful aspects of product development, leveraging AI as a powerful tool rather than a replacement.

The transformative potential of AI agents extends beyond software development, impacting the entire app market. Steinberger predicts that many existing apps may become obsolete as agents gain the ability to perform their functions more efficiently and contextually. This could lead to a significant shift in the software economy, with companies needing to adapt by becoming more API-centric or agent-facing.

While this transformation may cause short-term disruption, it also promises new opportunities and a more integrated, personalized digital experience.