![\"\"](\"https:\/\/vertu-website-oss.vertu.com\/2026\/01\/Clawdbot.png\")
<\/h1>\n<\/h1>\nClawdBot has swept Silicon Valley with 20,800+ GitHub stars and driven Mac mini sales to unprecedented levels, but the real story is even more remarkable: creator Peter Steinberger built this revolutionary personal AI assistant entirely solo, with 100% of the code written by AI\u2014not a single line hand-typed. In a candid “Open Source Friday” interview, the retired entrepreneur-turned-AI-builder revealed that ClawdBot started as a crude one-hour hack connecting WhatsApp to Claude Code, evolved through real-world testing (including a Morocco trip where it autonomously migrated itself between computers), and deliberately keeps 0.00001% closed-source\u2014a file called “soul” containing the AI's core values and identity\u2014as both secret asset and penetration test target that hackers worldwide are trying to extract (unsuccessfully so far). The project's radical collaboration model accepts PRs as “problem statements” rather than polished code, enabling non-programmers to contribute for the first time. Peter works 4 hours of sleep nightly, writes more code daily than his former 70-person company produced monthly, and insists ClawdBot represents “the year of personal agents” where everyone will have autonomous AI assistants\u2014but warns users: start with sandbox mode, use strong models (Opus recommended), never deploy on work computers, and understand the massive responsibility that comes with granting an AI complete system access.<\/p>\n
Peter Steinberger's Background<\/strong>: Ran successful B2B company for 13 years building world-leading PDF framework, team scaled to ~70 people, received irresistible acquisition offer<\/p>\n The “Retirement” Reality<\/strong>: Not traditional retirement but severe burnout after working weekends throughout 13-year journey<\/p>\n Recovery Period<\/strong>: Took time adjusting mentally and physically, compensating for life gaps, experiencing missed opportunities<\/p>\n The Return<\/strong>: Knew he was someone who loved “creating” and “building”\u2014inevitable return was just matter of time<\/p>\n The AI Catalyst<\/strong>: Early 2025, AI transitioned from “this doesn't really work” to “wait, this is interesting”<\/p>\n Evangelism Mode<\/strong>: Pulled countless people into AI exploration since that moment<\/p>\n Initial Motivation<\/strong> (April 2024): Wanted a “life assistant” but models weren't good enough yet<\/p>\n The Waiting Game<\/strong>: Put idea aside thinking “big companies will obviously build this\u2014why bother?”<\/p>\n The Realization<\/strong> (November 2024): Suddenly realized nobody had actually done it yet<\/p>\n The First Prototype<\/strong>: One hour of crude coding connecting WhatsApp messages \u2192 Claude Code \u2192 response back<\/p>\n Technical Description<\/strong>: “Essentially gluing a few things together\u2014honestly not difficult, but the effect was quite good”<\/p>\n The Image Requirement<\/strong>: Realized need for image input\u2014often used images when prompting because they provide agents with extensive context quickly<\/p>\n Morocco Testing Ground<\/strong>: Attended friend's birthday trip in Marrakech, used primitive system as “tour guide” while exploring city<\/p>\n Already Exceeded Expectations<\/strong>: Worked better than anticipated even in rough state<\/p>\n The Accidental Discovery<\/strong>: Sent voice message without implementing voice support<\/p>\n The Waiting<\/strong>: Watched “typing…” indicator wondering what would happen<\/p>\n The Shock<\/strong>: Few seconds later, it responded correctly<\/p>\n The Investigation<\/strong>: “What did you just do?”<\/p>\n The Autonomous Solution<\/strong>:<\/p>\n Host's Reaction<\/strong>: “Sounds like your first line of code triggered AGI”<\/p>\n Peter's Realization<\/strong>: “Maybe not quite AGI, but that moment I truly realized these things' ‘spontaneous adaptability' exceeded my original imagination”<\/p>\n The Hotel Lock Joke<\/strong>: Peter joked about unreliable Marrakech hotel door lock, hoped MacBook Pro wouldn't get stolen<\/p>\n The Agent's Response<\/strong>: “Don't worry, I'm your agent”<\/p>\n The Autonomous Migration<\/strong>:<\/p>\n Peter's Reaction<\/strong>: “I was thinking, this is the starting point of Skynet”<\/p>\n The Fundamental Truth<\/strong>: Not a single line of code in ClawdBot was manually typed by Peter<\/p>\n The Development Process<\/strong>:<\/p>\n Code Review Philosophy<\/strong>:<\/p>\n What Peter Doesn't Care About<\/strong>:<\/p>\n What Peter Scrutinizes Carefully<\/strong>:<\/p>\n The Missing Ingredient<\/strong>: “These agents still lack something: vision, taste, and love”<\/p>\n Against Meme Culture<\/strong>: Rejects notion of writing long requirements list and having agent complete everything automatically<\/p>\n Peter's Process<\/strong>:<\/p>\n Current vs. Original Understanding<\/strong>: “My understanding of this product is completely different from the beginning; in another month, after seeing how more people use it, it might change again”<\/p>\n Peter's Assessment<\/strong>: “In today's pace, the code I write in one day might exceed what my former 70-person company wrote in a month”<\/p>\n The New World<\/strong>: “In this new world, the speed of building things has completely changed”<\/p>\n Traditional PR Model<\/strong>:<\/p>\n That Made Sense When<\/strong>: Code was expensive and difficult to write<\/p>\n Current Reality<\/strong>: Code has become cheap\u2014feedback loop itself no longer valuable<\/p>\n New PR Philosophy<\/strong>: “PR is more like saying: ‘There's a problem here, this is my attempt to solve it.' I care more about what pain point this person really wants to solve, rather than whether the code is beautifully written”<\/p>\n Handling Approaches<\/strong>:<\/p>\n The Hard Part<\/strong>: Never writing code\u2014it's embedding new features reasonably into existing system<\/p>\n Architecture Knowledge<\/strong>: If unfamiliar with overall architecture, forcing features in causes eventual problems<\/p>\n PR as Problem Clues<\/strong>: “I'd rather treat PR as ‘problem clues' rather than ‘finished code,' otherwise project will slowly consume itself”<\/p>\n What's Open Source<\/strong>: 99.99999% of ClawdBot codebase<\/p>\n What's Closed<\/strong>: One file named “soul”<\/p>\n File Contents<\/strong>:<\/p>\n Dual Purpose<\/strong>:<\/p>\n Purpose 1: Secret Asset<\/strong>: Peter's 0.00001% proprietary advantage<\/p>\n Purpose 2: Security Target<\/strong>: Deliberately left as penetration test objective<\/p>\n The Challenge<\/strong>: Hackers worldwide attempting to extract “soul” file<\/p>\n Current Status<\/strong>: Nobody has successfully stolen it yet<\/p>\n What This Proves<\/strong>: Latest generation models show real improvement in prompt injection mitigation<\/p>\n Historical Context<\/strong>: “If you used very small, very old models and asked enough times, it would eventually say ‘okay, here's everything'\u2014that was our previous state”<\/p>\n Current Confidence<\/strong>: “Now with latest generation models, I'm confident: you must work very, very hard to possibly extract it”<\/p>\n The Slack Test<\/strong>: People constantly trying to hack Peter's agent in project Slack<\/p>\n Sandbox Caveat<\/strong>: Still not advisable to connect without sandbox to real environment\u2014demo agents run with limited permissions<\/p>\n Most Reliable Performance<\/strong>: Anthropic's Opus shows consistently stable results<\/p>\n Recommended for Production<\/strong>: Peter's primary suggestion for serious deployments<\/p>\n Why It Matters<\/strong>: Security and reliability critical when granting high system permissions<\/p>\n Surprising Winner<\/strong>: Among tested models, MiniMax 2.1 most “agentic” in behavior<\/p>\n Community Nickname<\/strong>: Internal team channel gave it playful name<\/p>\n Company Response<\/strong>: MiniMax embraced the meme, posted tweet saying “We may not have T0-level pricing, may not have team-level pricing, but at least we have target quality”<\/p>\n The Tweet<\/strong>: Became minor viral hit<\/p>\n Peter's Appreciation<\/strong>: “I really appreciate companies that don't put themselves on a pedestal too high”<\/p>\n Acknowledgment of Gap<\/strong>: “They're very clear they haven't caught up with top US labs technically yet, but in my view it's only a matter of time”<\/p>\n Acceleration<\/strong>: “Now many companies are accelerating catch-up, which itself is very exciting”<\/p>\n Local Deployment Advantage<\/strong>:<\/p>\n Peter's Enthusiasm<\/strong>: “This feeling is very cool\u2014honestly, almost no company can really achieve this”<\/p>\n Host Question<\/strong>: “Have you used Gemini model with ClawdBot? How's the experience?”<\/p>\n Peter's Response<\/strong>: “Gemini doesn't work now, really doesn't work well”<\/p>\n Host's Reaction<\/strong>: “Okay, conclusion very clear (laughs)”<\/p>\n Peter's Clarification<\/strong>: Believes tuning might extract better performance, but overall for tool calling and genuine “assistant feel,” hasn't found particularly good performance<\/p>\n Coding vs. Assistance<\/strong>: “Writing code is okay, but that's not this project's core focus”<\/p>\n Time Constraints<\/strong>: “Problem is, I only have so many hours in a day. I sleep four hours daily, rest of time writing code, haven't had time to polish everything yet”<\/p>\n Why Everyone's Buying Mac Mini<\/strong>:<\/p>\n High-Profile Adopters<\/strong>: Google DeepMind Product Manager Logan Kilpatrick ordered one<\/p>\n Extreme Cases<\/strong>: Users posting about buying 40 Mac minis simultaneously for ClawdBot deployment<\/p>\n The Reality Check<\/strong>: Can run identical program on free server<\/p>\n Alex Finn's Perspective<\/strong>: “No need to spend $600 on Mac mini\u2014there are much cheaper ways to run ClawdBot”<\/p>\n The Truth<\/strong>: Mac mini more about personal preference than technical necessity<\/p>\n Alternative<\/strong>: Don't need to buy any hardware\u2014just get a VPS<\/p>\n Most Comfortable Approach<\/strong>: “If you have an old computer at home”<\/p>\n Direct Usage<\/strong>: “Just use it directly”<\/p>\n Common Scenario<\/strong>: “Many people have an old Mac at home\u2014this scenario is simply perfect”<\/p>\n Installation Process<\/strong>: Website has single command, copy to terminal, step-by-step guidance follows<\/p>\n Cost-Effective Cloud Options<\/strong>:<\/p>\n New Solution<\/strong>: Install Gateway service in cloud, run node on local machine, connect securely via Tailscale<\/p>\n Hybrid Benefits<\/strong>:<\/p>\n The Aesthetic<\/strong>: “The most ‘flavorful' is still that old Mac. Someone put stickers on it saying ‘this is Claude's computer'\u2014I really love that image”<\/p>\n Windows Support<\/strong>: Also works, “just not as perfect, after all I have limited time”<\/p>\n Community Involvement<\/strong>: Pulled in contributors, looking for more to join<\/p>\n Peter's Most Extreme Integration<\/strong>: Connected ClawdBot to his bed<\/p>\n Eight Sleep Integration<\/strong>:<\/p>\n Home Automation Capabilities<\/strong>:<\/p>\n Communication Access<\/strong>:<\/p>\n The Trade-Off<\/strong>: “This is essentially a choice\u2014more permissions you give it, more powerful things it can do”<\/p>\n Twitter Automation<\/strong>: Favorite a tweet \u2192 automatically researches, organizes into to-do list<\/p>\n Full Application Development<\/strong>: People building complete apps with ClawdBot<\/p>\n Universal Mac Deployment<\/strong>: “Almost everyone configures it with a MacBook”<\/p>\n The 10,000 Email Purge<\/strong>: Peter's former business partner used it to clear inbox of 10,000 emails<\/p>\n Gmail Safety Net<\/strong>: “Clearing inbox” actually just archives\u2014doesn't truly delete<\/p>\n Multi-Person Setup<\/strong>: Each family member has own agent<\/p>\n Inter-Agent Communication<\/strong>: Agents can communicate and synchronize with each other<\/p>\n Shared Task Management<\/strong>: Family to-do lists automatically aligned<\/p>\n Peter's Admission<\/strong>: “Okay, I actually don't have a wife (laughs), but you can configure an agent for each person”<\/p>\n Coordination<\/strong>: “Your agent talks to my agent” replacing “let your people talk to my people”<\/p>\n Shopping Lists<\/strong>: Users generating grocery lists automatically<\/p>\n Tesco Integration<\/strong>: Some users have agents that can directly order from Tesco<\/p>\n Process<\/strong>: Say “buy these things again” \u2192 agent handles everything \u2192 items delivered to doorstep hours later<\/p>\n Pain Point<\/strong>: Host admits having expense report delayed one week, “really hate doing this”<\/p>\nThe Origin Story: A One-Hour Hack That Changed Everything<\/h3>\n
The Voice Message That Changed Everything<\/h3>\n
\n
The Skynet Moment<\/h3>\n
\n
Part II: The 100% AI-Written Reality<\/h2>\n
No Hand-Typed Code<\/h3>\n
\n
\n
\n
\n
The Production Speed Revolution<\/h3>\n
\n
\n
Part III: The Soul File Mystery<\/h2>\n
The 0.00001% Secret<\/h3>\n
\n
Part IV: Model Performance Rankings<\/h2>\n
Opus: The Stable Champion<\/h3>\n
MiniMax 2.1: The Most “Agentic”<\/h3>\n
\n
Gemini: The Disappointing Performer<\/h3>\n
Part V: The Mac Mini Phenomenon Explained<\/h2>\n
The Unexpected Hardware Star<\/h3>\n
\n
Peter's Recommendation: Use Old Computers<\/h3>\n
\n
\n
Part VI: Shocking Use Cases<\/h2>\n
The Bed That Obeys Commands<\/h3>\n
\n
\n
\n
Other Wild Applications<\/h3>\n
Family Agent Ecosystem<\/h3>\n
Expense Reports and Invoices<\/h3>\n