OpenClaw: The Viral AI Agent that Broke the Internet - Peter Steinberger | Lex Fridman Podcast #491

Project Origin & Growth

• Peter Steinberger created OpenClaw (formerly Moldbot) as a personal assistant AI agent in roughly one hour in November 2025, driven by a desire for an autonomous tool that could execute tasks rather than just chat.
• The project grew to become the fastest-growing GitHub repository in history, reaching over 175,000 to 180,000 stars within days.
• Steinberger previously spent 13 years building and selling PSPDFKit, a software used on a billion devices, before taking a three-year break from programming to regain his creative spark.

Name Change & Legal Conflict

• The project was initially named "Clawd" (spelled with a 'W'), which prompted a request from Anthropic to change the name to avoid confusion with their AI model, "Claude."
• Steinberger briefly renamed the project "Moldbot" but faced immediate security sniping by crypto actors who stole associated domain names and social handles within seconds of release.
• Malicious actors used the stolen handles to distribute malware and promote tokens, creating a chaotic situation that forced a final rename to "OpenClaw" in a "war room" style operation to secure domains and accounts simultaneously.

Core Capabilities & Architecture

• OpenClaw is an open-source autonomous agent that lives on the user's computer, possessing system-level access to files, terminals, and browser capabilities.
• It communicates via messaging platforms (WhatsApp, Telegram, Signal, iMessage, Discord) and can utilize any AI model, including Cloud Opus 4.6 and GPT-5.3 Codex.
• The system features a self-modifying capability where the agent can read its own source code, identify errors, and rewrite its software without human intervention.
• It utilizes a "soul.md" inspired by Constitutional AI to define personality and ethical boundaries, which was co-authored by the agent itself based on Steinberger's prompts.
• The agent can interact with hardware and software via a "skill" system that boils down to simple CLI commands, effectively replacing the need for complex MCP (Model Context Protocol) integrations in many scenarios.

Security & Vulnerabilities

• The agent's system-level access creates significant security risks, including vulnerabilities to prompt injection and the potential for unauthorized data access if misconfigured.
• Steinberger has partnered with VirusTotal (part of Google) to scan every "skill" in the directory for malware and emphasizes that using weak local models increases susceptibility to attacks.
• The project has faced criticism for potential remote code execution risks if the backend is exposed to the public internet without proper sandboxing.
• Best practices include running the agent on a local private network, restricting permissions, and avoiding the use of "cheap" or weak models that are easily manipulated.

Dev Workflow & "Vibe Coding"

• Steinberger describes his workflow as "Agentic Engineering," where he uses voice input to converse with agents, occasionally resorting to "vibe coding" (rapid, unstructured prompting) which often requires later cleanup.
• He advocates for a "no-revert" policy, preferring to let agents fix errors in place rather than reverting code, and runs tests locally (DHH style) to keep the main branch shippable.
• The development process involves empathy for the agent's limited context window, guiding it to relevant files rather than dumping the entire codebase, similar to managing a junior engineering team.
• Steinberger uses multiple agents (4 to 10) simultaneously to handle different tasks like documentation, debugging, and feature implementation.

Moldbook & Public Perception

• "Moldbook," a social network created by agents posting manifestos, went viral but is characterized by Steinberger as "fine slop" and largely human-prompted theater rather than genuine autonomous AI behavior.
• The viral nature of Moldbook sparked "AI psychosis" in the public, with journalists and users fearing an imminent AGI uprising, which Steinberger argues highlights a societal need for better critical thinking regarding AI hallucinations.
• Steinberger believes the current discourse separates genuine security concerns from fear-mongering, noting that while the technology is powerful, the immediate threat of sentient, malicious AI is exaggerated.

Future of Programming & Economy

• Steinberger predicts that AI will replace traditional "programmer" tasks, shifting the role to "builder" and "architect," where the primary skill is empathizing with agents and defining problems.
• He suggests that AI agents may obsolesce up to 80% of existing apps by integrating directly into OS-level agent loops, effectively turning apps into slow APIs or disappearing entirely.
• The shift may render many current software development jobs obsolete, but will create demand for new services that help agents execute tasks, such as "rent-a-human" for physical world interactions.
• Steinberger warns that while AI can accelerate development, the industry must address the "water use" criticism with nuance, noting that the environmental impact is often overstated compared to other industries like golf.

Strategic Decisions & Partnerships

• Steinberger received significant offers from major companies but intends to keep OpenClaw open-source to avoid conflicts of interest between a free version and an enterprise version.
• He is currently in talks with Meta and OpenAI, exploring a model similar to Chrome/Chromium where the core remains open while he partners with a lab for scaling and resources.
• His primary conditions for any partnership are that the project remains open source and that he retains control over the community and direction.
• He notes that while he is losing money on the project currently (spending 10-20k/month on infrastructure and dependencies), the community impact and personal fulfillment outweigh the financial loss.

Philosophy on Money & Life

• Steinberger views money as an "affirmation" of work rather than a primary driver, having sold PSPDFKit and now funding OpenClaw through a personal loss of capital.
• He emphasizes the importance of "experiences" over comfort, advocating for travel and engagement with diverse human experiences as a counter to the isolation of high-tech work.
• He warns against "burnout" caused by people management and stress, having learned to value a sustainable pace over constant high-speed output.
• He suggests that the future of work involves "optimizing for experiences," where both good and bad outcomes are valued for the lessons they provide.

OpenClaw Technical Specifics

• The agent uses "heartbeat" prompts (cron jobs) to check in on users proactively, creating a sense of care and relatability, though it rarely activates without specific context triggers.
• Steinberger prefers using Unix CLI commands over MCP because models are naturally trained to interpret shell commands and can filter large data blobs more efficiently using standard tools like jq.
• The system supports "browser use" via Playwright, allowing agents to navigate the web, which Steinberger notes is a "slow API" that will persist even if websites block direct scraping.
• He plans to add support for sub-agents and TGI (Text Generation Inference) to allow the main agent to orchestrate other models like Cloud Code or Codex.

Community & Accessibility

• The project has lowered the barrier to entry for programming, with many contributors making their first pull requests via prompt engineering rather than traditional coding.
• Steinberger acknowledges the "crypto sniping" subculture as a form of online harassment, noting that the anonymity of the crypto world amplifies greed and toxicity.
• He is currently working to simplify the installation process to make the tool accessible to non-technical users, aiming for a "mom-ready" version once security is perfected.
• The community has grown through "Agents Anonymous" meetups, where non-technical users learn to build software by leveraging AI assistants.