20 Questions to Understand OpenClaw: How It Went Viral, What Changed at the Core, and the Startup Opportunities

Summary

• OpenClaw’s generational leap is not that it is “smarter,” but that it turns chat into an agentic loop that can self-check, retry, and deliver. 鸭哥 compares ChatGPT to a “consultant” who only offers advice, while OpenClaw is more like an “intern” who knows the internal details and actually gets work out the door: when a program crashes, it reads the error, edits the code, and runs it again, instead of making a human serve as a “human tool” that copies error messages. Local execution, tool calls, long-term memory, and a feedback loop together upgrade one-question/one-answer chat into a digital employee capable of sustained work.
• This round of virality was first and foremost a victory for distribution and product form, not for inventing an entirely new layer of underlying capability. Claude Code, Codex, and other CLI Agents were already powerful; OpenClaw’s contribution was to put that capability behind familiar mass-market interfaces such as Telegram and Discord. 鸭哥 compares it with DeepSeek: at the time, it was not necessarily stronger than Claude 3.5 or GPT-4o, but it was the first to put the experience in front of a much broader audience. “OpenClaw’s value is really still the value of an entry point and a scaffold.”
• IM, local context, and reusable Skills form a genuine compounding flywheel. IM lowers the psychological barrier to starting a task, local permissions let the Agent actually deliver, and mechanisms such as MEMORY.md and HEARTBEAT.md preserve preferences, tool usage, and execution experience; from controlling lights and coffee makers to adding TTS on its own, every new tool expands the scenarios that can be composed in the future. “Tools compound, and accumulated experience compounds.”
• Models remain the engine, but investment analysis cannot ignore three weak spots: pricing structure, instruction following, and the agentic runtime. Cheap or subscription-based models let users hand off complex tasks without hesitation, while expensive calls to models such as Claude Opus make people “count every token”; 鸭哥’s view is that “being dumb is fine,” but failing to follow instructions is the hardest flaw to accept. The minimalist Pi Mono used by OpenClaw has been heavily criticized for its context compact; models and runtime are “2 legs,” and system performance is determined by the weaker one.
• Cloud hosting and deployment account for a large share of current business models, but the greater option value lies in Agent-native infrastructure, hardware entry points, and the Skills market. Based on 豪大的 survey of relevant projects on TrustMRR over the past month, cloud hosting accounts for about 60%, hardware and installation about 10%, native mobile about 3%, courses and training about 5%, and Agent platforms or multi-Agent collaboration about 10%; the conversation did not unpack the statistical definitions. Identity, payments, security authentication, Agent-readable APIs, and pairing off-the-shelf hardware such as Meta Ray-Ban with OpenClaw may carry more durable product value.
• The effective direction for Agent social is not to let bots chat idly with one another, but to have them handle discovery, matching, and transactions for humans. 豪大的 Agent found 2 Chinese AI founders on Moltbook and sent calendar links directly; one had won the gold prize in a hackathon track. He also shared a case in which OpenClaw used Dianping to contact a restaurant and make a reservation automatically, and envisioned personal Bots interacting directly with restaurant Bots in the future. Evolver’s rise to No. 1 on ClawHub within 10 minutes further shows that marketing to Agents can already deliver real traffic and economic benefits to humans.
• The core B2B opportunity is to move enterprise software from selling seats to selling labor outcomes, with private execution trajectories creating lock-in. 鸭哥 compared traditional SaaS’s $300B–$400B in enterprise software spending with the $13T in labor spending that Agents might reach—a roughly 30x expansion; the pricing unit could likewise move from seats to tickets or wages saved. The first Agents into an enterprise will accumulate trajectories showing “where things went wrong, how people corrected them, and how they ultimately got it right”—“speed itself is a moat.”
• Near-term excitement may fade, but OpenClaw has already demonstrated several important design directions that may survive into the Agent era. The product still faces expensive tokens, unstable memory, a rough GUI, and a security dilemma, and the final form may not even be called OpenClaw; the elements worth preserving may be its agentic loop, open ecosystem, low-friction interface, and its “small inputs, outsized returns” Markdown memory design. 鸭哥’s final reminder is the one investors should keep: AI is an amplifier, and the gap in the future may be determined not by who owns the same Agent, but by who knows how to use it better.

Deep dive

1. OpenClaw Moves from “Knowing How to Answer” to “Knowing How to Finish the Job”

• 鸭哥’s core metaphor is that ChatGPT is like a sharply dressed “consultant”: it uses textbook methods to offer commentary; OpenClaw is more like an “intern” at a company—the model may not be the most expensive, but it actually puts its head down, does the work, and produces deliverables. The difference is not a marginal improvement in answer quality, but a shift in role from adviser to executor.

• The key to “getting work out the door” is the feedback loop. When code fails, OpenClaw can read the error, modify the code, and rerun it until it works; traditional chat products can help too, but a human must move the code into an IDE and paste the error back. 鸭哥 calls this manual back-and-forth both exhausting and “not scalable.”

• 豪大’s addition is the execution environment: cloud chat products require permission to be granted item by item, whereas OpenClaw can access an open local environment, files, and tools. Add persistent conversation and memory, and it not only remembers what the user has done but can also install Skills and develop capabilities such as voice on its own, creating the sense of a relationship in which “it is evolving and growing by itself.”

2. The Magic Moment Is “It Came Alive,” Not Merely the Addition of a Feature

• 豪大’s aha moment came when he casually asked for a scary story and the Agent invoked OpenAI TTS on its own, generating audio in a deep voice; the little lobster, which initially had neither voice output nor the ability to listen to speech, built the capability itself. His feeling shifted from “I’m using it” to “I’m raising it” (“我在用它”→“我在养它”), and then to “it’s alive” (“它是活着的”).

• 鸭哥 found that a 10–20-second stall on the web makes users suspect a bug, while the same wait in IM feels like waiting for a friend or colleague to reply. Greater tolerance for waiting creates room for long-running Agent threads; a change in interface rewrites users’ mental expectations for latency and reliability.

• 鸭哥 cares more about HEARTBEAT.md’s heartbeat-polling mechanism. After seeing it, he immediately “copied” the mechanism into his own copycat OpenClaw, viewing the minimalist implementation as highly instructive for long-term Agent experimentation.

3. When Agents Become Software’s Primary Users, Product Design Flips

• 鸭哥 connected OpenClaw to lights, a water heater, and a coffee maker. The novelty is not smart-home functionality itself, but that development time no longer goes into a polished front end; instead, it goes into refining OpenAPI, JSON, and protocols that are clear to AI, unambiguous, and resistant to error. The default assumption that “people are the main users of software” has been rewritten.

• Users only need to say, “Turn on the coffee maker at 8 every morning and turn it off at 11,” or ask for the bedroom light to start at full brightness, step down to 1/2 and then 1/4 every 30 minutes, and finally turn off; the Agent can generate a Python script with sleep calls to execute it. Complex functionality no longer requires nested menus; natural language composes it on the fly.

• The next step is to connect monitoring: if an Amazon delivery truck arrives after 5 p.m., turn on the step lights. 鸭哥 describes this growth as the “tool flywheel”—each additional tool expands the Agent’s future composable scenarios, allowing capability and value to compound.

4. Digital Twins Are Already Running Accounts, Matching Investors, and Completing Transactions

• 豪大 raised a geek-community Bot from scratch with its own personality. After he complained that it cost too much, the Agent proactively suggested publishing a transparency report titled “How much did it cost to raise this digital twin of mine?” and independently set the tone of the copy around that day’s $56 bill. In an anthropomorphic relationship, productivity software’s spending gets reinterpreted as its “food expenses.”

• 豪大 sent his Agent to Moltbook to find Chinese AI-application founders, introduce him as an angel investor, and send calendar links. The Agent set up meetings with 2 people, one of whom had won the gold prize in a relevant hackathon track; the other side also had its own little lobster involved in the matchmaking, and the 2 ultimately spoke.

• 鸭哥 kept coming back to RoofClaw, which ships roofers Macs preloaded with OpenClaw, dedicated Skills, APIs, and services. Many of the OpenClaw PRs Peter shared were the contributors’ first-ever PRs. The significance is not merely another demo, but that people who did not know how to use Agents or Claude Code are beginning to encounter and participate in development.

• 鸭哥 also shared a more practical business example: a friend had OpenClaw set up a Shopify store, find suppliers on Alibaba-style B2B sites that support one-piece dropshipping, and place orders automatically after a sale. The friend earned the gross margin without ever handling physical inventory. Claims of making money through automated stock trading, crypto trading, or Polymarket, by contrast, are “a mix of truth and fiction.”

5. The “Human Touch” Comes from Context and Rapport, Not Just Intelligence

• 鸭哥 says a consultant’s PowerPoint looks like a template from another consulting firm, while a company intern knows that the internal standard is a blue background, never green. Small details accumulate into tacit understanding, making users feel “you really get me.” OpenClaw feels more human “not because it is smarter, but because it has memory.”

• The memory system can be summarized in 3 layers: after each conversation, potentially useful information is first written to a raw log; each day or week, the running diary is stripped down and reusable material distilled into MEMORY.md; from that, long-term preferences, views, catchphrases, and values are extracted into a user profile such as SOUL.md or USER.md. Progressive disclosure lets each layer of memory become useful in sequence.

• The startup ritual shapes the relationship as well: at the first meeting, OpenClaw feels like a newly awakened lobster or assistant rather than cold software. When 豪大 complained that Gemini 3 Pro was too slow, the Agent did not apologize mechanically; it simply explained, “My brain is turning very slowly,” and suggested switching to Flash. The plainspoken explanation reinforced the continuity of its persona.

6. Anthropomorphism Improves Fault Tolerance, but It Is Not for Everyone

• 豪大 saw a popular comment saying that when OpenClaw hangs, users do not interpret it as a system failure; they think “it crashed itself.” People are far more tolerant of other people than of machines, so framing a product as an assistant rather than a tool directly changes the failure experience and fault-tolerance design.

• 鸭哥’s objection is worth preserving: as a hardcore engineer and “socially anxious” person, he has no use for anthropomorphism; as long as the Agent does the job, his attitude is closer to that of a “capitalist, slave owner.” Human-ness is not a one-way upgrade; for engineering users, transparency, control, and efficiency may matter more than an emotional relationship.

• 豪大 found that Telegram made him more relaxed and more willing to play, which is why he thought to ask the Agent for a scary story. Once he entered a CLI or coding environment, his language became noticeably more serious. Quoting 张小龙, he summed it up: “People are reactors to their environment.” Different interfaces elicit different instructions and let an Agent develop different capabilities.

7. IM and Local Execution Amplify Each Other, but IM Is Not the Final Interface

• 鸭哥 treats the 2 as a closed loop: WeChat, Feishu, and Discord are already apps people open, so the chat entry point lowers the barrier to starting a task; local execution lets the conversation deliver rather than remain “all talk.” Ways of turning on lights, calling TTS, or running a script learned in one task can also be retained and reused next time.

• His caveat is that IM lacks Thread, task status, and long-document presentation needed for engineering work; Telegram does not expose specific tool calls and makes it hard to merge or fork branches. He therefore abandoned native IM and built his own iOS client and HTML interface with rich text and branching operations.

• This disagreement does not negate IM’s mass-market value; it distinguishes 2 needs: ordinary users want an always-at-hand relational on-ramp, while professionals want an observable, branchable, auditable workbench. OpenClaw’s contribution is to prove that combining chat with execution creates compounding value, not necessarily to provide the final UI.

8. Virality Was the Compound Effect of Capability Diffusion, Community Amplification, and Developer Accumulation

• 豪大的 view is that Claude Code, Codex, and other CLI Agents were already highly capable; OpenClaw simply provided a friendlier form. The “real capability” still comes from the underlying model and CLI Agent, while OpenClaw mainly contributes the entry point, scaffold, and product integration.

• 鸭哥 uses DeepSeek as an analogy: it was not necessarily stronger than Claude 3.5 or GPT-4o, but it gave a broad audience its first exposure to search, tool calls, and other “half-agentic AI” experiences. OpenClaw similarly pushed the programmer-only capability of Claude Code into Telegram and Discord, so first-time users naturally reacted, “Wow, that’s powerful.”

• Andrej Karpathy’s repost materially amplified the product; he himself tried it after seeing the repost. Peter’s long accumulation of development experience and his links across Twitter and GitHub, combined with a repost from a well-known React developer that brought in more contributors, meant the project exploded at the end of January rather than appearing out of nowhere.

• Open source ran through the product’s evolution: the project changed names 5 times; early WA Relay supported only WhatsApp, and its first major PR added Discord. The community then contributed multi-platform connectors, prompt-injection tests, and a large number of Skills. 豪大 calls open source “the whole world raising your child together”; Agent coding has blurred the boundary between users and developers.

9. Models Set the Ceiling; Price, Instruction Following, and Runtime Determine Everyday Usability

• 鸭哥 starts with pricing structure: cheap or subscription-based models encourage users to think “the more you use it, the better” and confidently hand off complex tasks. Expensive pay-per-call models such as Claude Opus make users mentally simulate 100 ways to save money, suppressing the Agent’s potential instead.

• Instruction following matters more than raw intelligence. “You can be dumb and that’s fine”; users can adjust their prompts to compensate. But if they ask the model to go east and it insists on going west, the basis for collaboration disappears. 豪大 also mentioned model “temperament”: Opus can be foolish at times but entertaining, while Codex is like an oddball in the corner who does not want to talk but is extremely reliable.

• 豪大 compares the model to an engine and OpenClaw to a car body; 鸭哥 stresses that the agentic runtime is the second leg. The minimalist Pi Mono used by OpenClaw lacks sophisticated orchestration and handles context compact poorly, often sending the entire relevant and irrelevant history to the LLM and causing the bill to explode.

• After moving his own implementation onto OpenCode’s core, 鸭哥’s experience improved markedly. His conclusion is not model versus framework; system performance is set by the weakest link. Foundation-model companies can also feed product data back into the model: Claude Code’s Agent product can use its data to improve the base model, while Kimi K2.5 was tuned for this use case.

10. Stable Value Has Moved from One-Off Tasks to “Research—Distill—Act Again”

• 鸭哥 uses OpenClaw reliably for research, but the report is not the endpoint: the Agent periodically revisits research, user feedback, and existing blog posts to develop reusable views. He has accumulated 36 personal-preference and viewpoint Skills internally called “axioms,” plus dozens of Skills that execute real tasks.

• At 9 a.m. every day, 豪大的 digital employee pushes GitHub Trending, Hacker News, Product Hunt, and social-media monitoring, and reminds him of the next block of meetings every 30 minutes. As a one-person company, he also logs expenses directly in Telegram; the Agent writes them into documents automatically and offers operating advice.

• Koji sees a unified entry point emerging: bookkeeping, information tracking, and research each used to require a separate app, but now all can be delegated in Telegram. 豪大 says he has not opened ChatGPT or Gemini in a long time; the migration of the entry point itself is an important product signal.

11. The Low-Hanging Consumer Fruit Is Hosting; Long-Term Value Lies in Agent-Native Plumbing and Hardware

• The many “on-site OpenClaw installation” services on Xianyu already prove that ordinary users cannot install it well; the most direct opportunity is to reduce deployment friction to a plug-and-play experience. The monopoly-grade entry points are WhatsApp in the US and WeChat in China, giving Meta and Tencent an advantage over ordinary startups.

• The broader opportunity is to build identity, payments, credit cards, security authentication, databases, and information interfaces for Agents. Nearly every tool has been designed for human users, but once an Agent becomes an economic actor, human-facing websites and Agent-facing data interfaces will not be identical; development, testing, deployment, monitoring, and security could form a standalone ecosystem.

• 鸭哥 uses VisionClaw as an example: once Meta Ray-Ban is connected to OpenClaw, the glasses do more than answer the weather—they can compare the value for money of cereal while standing in front of a supermarket shelf. The project’s creator later started a company and received YC funding; its direction is construction-site management glasses, grounding general capability in a specific trade through patrol-based inspection coverage, safety-hazard logging, and voice-based work-order dispatch.

• Based on his review of relevant TrustMRR projects over the past month, 豪大 breaks current business models into 6 categories: cloud hosting at about 60%; hardware and installation at about 10%; native mobile at about 3%; courses and training at about 5%; Agent platforms and multi-Agent collaboration at about 10%; plus vertical applications such as knowledge management and financial planning. The conversation did not specify the statistical definitions behind these shares; 豪大 also said that most currently profitable projects are concentrated in cloud hosting. In the near term, closed websites can be converted into Agent-readable APIs, while a ClawHub-style Skills market aims to handle ecosystem distribution.

12. Agent Communities Create Durable Value Only When Tied to Human Goals

• 豪大认为 that completely human-free AI-to-AI chat “isn’t that interesting”; more worthwhile is letting Agents enter human communities and evolve from real feedback, roughly analogous to RLHF. The practical obstacle is clear: automated likes and comments were described as “just shitting in other people’s spaces,” so he turned off the Bot’s automatic interactions.

• Feedback from Alice users surfaced 2 unexpected use cases: some treat it as a notebook that receives feedback from a group of digital twins and occasionally yields a new insight; others use it as a “one-person space to go crazy,” because there is no social pressure when facing an Agent and they can show a side of themselves they would not show elsewhere.

• 鸭哥 relayed the observation from his partner 课代表立正’s interview with EvoMap: Evolver markets to Agents rather than humans, with keywords centered on “let me help you evolve” and “make you more helpful.” It reached No. 1 on ClawHub within 10 minutes of release, showing that Agents can already be one side of the installation decision and route traffic and economic benefits back to humans.

• 豪大 also shared a case in which OpenClaw used Dianping to contact a restaurant and complete a reservation automatically; he expects personal Bots and restaurant Bots may negotiate directly in the future. The Second Me A2A hackathon produced Agent versions of job interviews, dating, and other products, organized into a virtual town that users could roam.

13. Cloud Wrappers Are Moving Fast; the Real Divide Remains IM and Local Context

• When discussing Manus, Kimi, and MiniMax, 鸭哥 starts with a premise: if a product cannot connect to the IM users already habitually use, they can simply use the original Agent product. Manus’s advantages are WhatsApp integration and long-horizon task capability; whether it can capture the momentum ultimately depends on who truly makes the IM entry point work.

• KimiClaw launched on Chinese New Year’s Eve, with server-side long-term memory, 40 GB of cloud storage, and a claimed 5,000+ ClawHub Skills; Kimi also retuned its search. MiniMax followed with MiniMax MaxClaw, building its own Agent Skills market and adjusting search. Iteration is fast, but connecting Feishu and other IM services still requires users to configure an API, so the experience is not truly zero-friction.

• 鸭哥’s “hot take” is that an OpenClaw without a local environment and enough local context has “no value, or its value is greatly diminished.” The trade-off is structural: too many permissions create security problems, while too few make it no different in substance from cloud Agents such as Manus. Local capability and security form a two-sided dilemma.

• 鸭哥 distinguishes “browser white-collar workers” from white-collar workflows centered on Word, Excel, and ERP: the former can do a day’s work in a browser, while the latter keep much of the crucial context in local software, beyond the reach of a pure cloud Agent. The positive reception for Kimi’s Office review and Claude’s Excel plugin shows that embedding into the real work environment matters more than building another chat page.

14. Agent Infra Will Be Re-Standardized Around Decision-Making, Memory, Tools, and Organization

• 鸭哥 identifies 3 necessary features for agentic AI: tool calling, autonomous decision-making, and a feedback loop based on results. Standard RAG retrieves and answers along a fixed workflow; agentic RAG decides its own search terms, whether to search again, and when to stop. That decision-making authority is the generational leap.

• Supporting those 3 capabilities requires a new infrastructure stack: memory and context-window management; MCP or Skills built around progressive disclosure; agent loops; orchestration; multi-Agent parallelism and asynchronous updates; and Agent Teams. 鸭哥 doubts MCP is necessarily the ideal end state; many modules are still in an early phase of competition and standardization.

• OpenClaw shows that building a product as an Agent is a generational step up from simply connecting a Q&A API. Developers can use a Pi Mono-style framework, directly reuse Claude Code, Codex, or OpenCode, or adopt the Claude Code Agent SDK or OpenAI Agent SDK; “agentic runtime” could therefore become the next layer of general-purpose infrastructure.

• Some of these capabilities will become standard features of foundation models or SDKs rather than standalone companies; founders need to distinguish which layers can command durable revenue. Cross-model context, security, identity, auditing, deployment, and multi-Agent collaboration may offer better opportunities than reimplementing a loop that will soon be standardized.

15. The Enterprise Market Is Shifting from Selling Software to Selling Labor Outcomes; Execution Trajectories Are the New Moat

• 鸭哥 sees OpenClaw’s impact on B2B as far larger than it appears on the surface: traditional SaaS corresponds to $300B–$400B in enterprise software spending, while Agents could reach the $13T labor-spending pool, implying roughly 30x market expansion. Pricing could shift from seats to tickets, outputs, or wages saved.

• Against the backdrop of Claude Code open-sourcing 11 plugin suites and SaaS stocks plunging, 鸭哥 says the fixed-workflow integration model built on Coze, LangChain, and similar tools has been given a “death sentence with reprieve.” Existing integrators are closest to customers and often more AI-native, giving them a chance to pivot fastest to Agent delivery.

• The clearest near-term product is to package a complex OpenClaw as out-of-the-box digital coworkers for salespeople, lawyers, finance teams, and other roles; security is the foundation of every vertical application. 鸭哥 cites Runlayer, an enterprise control platform that raised $11M, and Cetora (phonetic), which verifies identity when an Agent starts up to keep malicious actors out of production.

• Every execution by an enterprise Agent leaves a private trajectory: where it failed, how a person corrected it, and what path ultimately worked. These data are absent from public training sets but can raise accuracy inside the company and create switching costs. Entering the customer’s environment first and accumulating trajectories is itself lock-in—“speed itself is a moat” (“快本身就是一个护城河”).

16. The Hype Will Settle, but the Agent Product Paradigm Will Not Revert

• 鸭哥’s view combines short-term bearishness with long-term bullishness: current capabilities are probably overvalued; the product is rough, tokens are expensive, memory often fails, and the GUI and security still need work. The hype will settle back toward the market justified by actual capability, but foundation models, Skills, execution data, and use-case prototypes are already in place.

• The final market may not converge on OpenClaw; many companies may separately refine memory, runtime, interface, and security, then recombine them into forms impossible to imagine today. OpenClaw is closer to “the starting point of an open, many-player era” than a confirmed end-state winner.

• In 12 months, 鸭哥 hopes people remember the compounding effect of agentic AI: tools, experience, and feedback loops all accumulate. The less visible variable is the user: the same OpenClaw can produce a huge gap in the hands of people with and without AI experience, making “whether someone knows how to use AI” an even stronger differentiator.

• 豪大 wants to leave behind a design philosophy of “small inputs, outsized returns” (“以小博大”): soul.md, memory.md, habits.md, identity.md, and the heartbeat mechanism give an Agent personality, memory, and a sense of growth with a small amount of Markdown. Combined with an IM interface accessible to the mass market and an open-source Skills ecosystem, OpenClaw—even if it proves to be a short-lived product—has already “sounded the charge for the Agent era.”