OpenClaw Journey: From WhatsApp Relay to Open‑Source AI Phenomenon

The speaker, known for creating OpenClaw, reflects on the journey of building and maintaining an open-source AI project, sharing insights on its rapid growth, the challenges faced, and lessons learned.

The Genesis of OpenClaw

The idea for OpenClaw originated from a personal annoyance. In November 2024, while juggling multiple AI agents and preparing food, the speaker realized there was no efficient way to send prompts from a phone to a computer to check on agent progress. This frustration, combined with the limitations of early 2025 AI models (like OpenAI's slow and expensive GPT-3), sparked the desire for a more seamless interaction.

The speaker's initial experiment involved building a WhatsApp relay using a coding agent. This allowed messages to be sent between a Mac and WhatsApp, which felt "magical" due to its proactive and concise responses, making the complexity of AI interaction melt away. The system was designed to feel like a "friend," prompting a belief that this was the future of AGI.

Despite the speaker's enthusiasm, others didn't initially grasp the significance. This annoyance led to creating group chats with friends, adding the WhatsApp relay, and observing their strong emotional reactions—ranging from amazement to fear. The demand from non-technical friends, who were upset they couldn't access it, signaled a strong product-market fit.

From Relay to OpenClaw: The Viral Explosion

After months of refinement, a pull request to add Discord support to the WhatsApp relay, despite the project's name, led to the creation of "Claudis" (later OpenClaw). This was a rudimentary system, lacking features like compaction.

Around New Year's Eve, the speaker dedicated more time to the project. In the first week of January, as the tech world was learning about coding agents, OpenClaw was launched in a Discord server. The initial night was a blur of activity, with people joining, trying to hack it, and interacting with the AI. The speaker had implemented a safety measure, instructing the agent not to perform dangerous tool calls unless prompted by "Peter."

At 7 AM, after an all-night session, the speaker went to bed, but the software, designed to be resilient, restarted due to being a launch daemon. Upon waking 10 hours later, the speaker found 800 messages, indicating the project had gone viral. This led to an explosion of attention: reporters calling at odd hours, a flooded email inbox, Mac Mini sales skyrocketing, and numerous podcast invitations.

Anthropic, a company, demanded a name change, leading to the project evolving from "Claudis" to "Claudebot" and eventually "OpenClaw." Jensen Huang of Nvidia reportedly called it "the most successful open-source project in the history of humanity."

The Overwhelming Success and Its Toll

The numbers were staggering: over 18,000 unique individuals opened an issue or pull request, with a total of over 111,000 interactions. Nearly 3,000 people contributed commits.

However, this success came with significant challenges. The speaker was unprepared for the intense attention, which almost led to deleting the entire project. Friends were ignored, and the phone became a source of dread due to constant notifications. Personal details were leaked, and the speaker became the target of both praise ("AI Jesus") and condemnation ("antichrist").

Lessons from a Burnout and a Comeback

The speaker had previously built and sold a B2B software company, experiencing burnout. After three years of aimless wandering, the urge to build returned, not necessarily to code, but to create.

When major AI labs like OpenAI and Anthropic came knocking, the speaker faced a critical decision. The experience highlighted the importance of a personal brand, as products can be forked, but a name cannot. The decision to engage with these companies, while feeling like "selling my soul," was guided by gut instinct, leading to a preference for OpenClaw.

The Challenges of Open Source and External Dependencies

OpenClaw faced significant security reports after its release. While most were "super hacky," the pressure was immense. A paper was published to correct the media's claim of 20% malicious skills, showing the actual figure was 0.3%. The speaker felt a strong responsibility to harden the codebase, implementing sandboxing, allow lists, and a web protocol with built-in permissions. However, these security measures often made the software slower and harder to update, frustrating users who prioritized functionality over security.

The speaker's time became divided between open source, dealing with the press, and setting up a 501(c)(3) nonprofit. The community of maintainers, working for free, added numerous features, leading to a complex system with around 9,500 configuration options. This made testing and evolving the software incredibly difficult.

Meanwhile, VC-funded companies, particularly the "anime girl company," pushed ahead with aggressive marketing and simpler solutions. A major blow came when Anthropic, whose model OpenClaw was heavily optimized for, announced they would disable subscriptions with only 24 hours' notice. This highlighted a crucial lesson: "Your dependencies' business model is your business model." While OpenClaw supported open-weight models, they weren't as good at the time, and early OpenAI models "lacked character."

Rediscovering the Fun and Future Directions

Around February, the project stopped being fun and became a responsibility. The speaker, who didn't want to build another company, found himself with two jobs. The product shifted from something personally loved to something built for "everyone," leading to a loss of personal connection. The speaker became primarily responsible for fixing bugs, security issues, and providing support.

In hindsight, the speaker acknowledges the need to have asked for more help and delegated responsibility. Support from Nvidia, who provided security personnel, was particularly helpful. By May, the joy of building began to return.

Despite numerous "OpenClaw killer" projects emerging, the speaker emphasizes that the core of OpenClaw is about open source and the joy of building. The current annoyance is having to use software that can't be easily modified with a prompt to an agent.

Looking ahead, OpenClaw's mission is to bring people closer to AI, making it fun and accessible. The foundation is growing, with 10 people on payroll and plans to hire a CEO. The term "claw" has become ubiquitous in the AI space, with 33,000 repositories bearing the name.

The speaker believes there's still much to do, particularly in creating always-on, always-syncing agents. The AI landscape is evolving rapidly, moving beyond text-only interfaces to voice and multimodality. OpenClaw aims to be the open-source alternative to proprietary agents, running everywhere, with any model, and keeping data local.

The project is now being built with OpenClaw itself, using a team server where agents can see each other's sessions and orchestrate work. This moves beyond the current paradigm of working in terminals with laptops open.

Key Takeaways

1. Don't stop having fun: Fun is the ultimate driver for generating the best ideas.
2. Listen to your gut and fix annoyances: This can lead to the next big thing.
3. Stay focused: Avoid distractions and build what's missing.

Q&A Highlights

• Building with agents: The speaker now focuses on making agents do more proactive work, expecting fully reviewed and tested pull requests rather than raw issues.
• What comes after graphs? The speaker views "loops," "graphs," and "workflows" as different terms for the same concept: designing automations that respond to triggers and perform actions.
• Reliability and scalability: The speaker admits to having "messed up" on this initially due to lack of focus and models not being good at testing. Modern models, with orchestration capabilities and tools like browser use, offer a much better QA environment.
• Early decisions for speed: The speaker learned to view code review as risk management, not necessarily reading every line, but focusing on areas of higher risk or unexpected complexity.
• Getting first users: The first user should always be the builder themselves, followed by friends. Excitement for the product is crucial.
• Balancing issues and features: This is a constant challenge. While fixing issues is important, a healthy mix with building new features is necessary to maintain interest.
• Changing one thing about OpenClaw: The speaker would have been less stressed about security researchers, recognizing that many were seeking attention rather than genuinely helping the product.
• Biggest bottleneck in AI infrastructure: Reliability, tooling, memory, and managing compute are major challenges.
• Keeping an open-source project opinionated: The speaker admits to not saying "no" enough to pull requests. A clear "vision.md" file is crucial, and every merged feature adds responsibility and complexity.
• Perpetually running, proactive claws: This is more a "token problem" than a tech problem. Designing systems that don't burn empty tokens and optimize for large sessions is key.
• Personal setup: The speaker uses a MacBook but primarily screen-shares into a studio computer for continuous operation and to prevent agents from interfering with the main machine.
• Next startup idea: Build something you want to use, work on personal brand and visibility, and consider "hard and boring" categories where solving problems is appreciated.
• Product to build: A reliable, fast, and cheap test box for Mac (and Windows) is a significant unmet need.