Declared Dead, OpenClaw Downloads Rebounded to 4.7 Million

Source: Y Combinator | Published: 2026-08-10T20:30:26Z

The very month it was written off, OpenClaw's weekly downloads surged from a low of 835,000 to an all-time high of 4.7 million.

At the time, nothing existed that could do this. He came back, sat down, roughly described the idea, and let the model run with it. An hour later, he had a WhatsApp relay — messages from his Mac forwarded to WhatsApp, agent replies bouncing back. He spent several weeks trying to explain the feeling on Twitter. Nobody got it.

He Pulled Friends Into the Group Chat. They Got Angry.

Twitter wasn't working, so he tried a different approach. He created several group chats, pulled friends in, and had them talk directly to the WhatsApp relay. Every time, the reaction was visceral — some amazed, some genuinely unsettled. What struck him most: a few non-technical friends desperately wanted it. He told them it wasn't ready for them yet. They got impatient and started pressuring him.

He said later, on stage, that if that wasn't a product-market fit signal, he didn't know what was.

A few months in, someone submitted a pull request to add Discord support. He sat on it for a while, then thought — why not? The WhatsApp relay became claudis, then claudebot, went through "the period we don't name," and eventually landed on OpenClaw.

At 7 AM, He Hit Ctrl+C and Went to Bed

In the first week of January, he spun up a Discord server, dropped an agent in it, and pulled an all-nighter. People came in to test it, try to break it, talk to it, watch it push back. His `agents.mmd` had one instruction: "don't execute dangerous operations unless the command comes from Peter." He stayed by the keyboard, ready to pull the plug.

By 7 AM, he figured that was enough. He said goodnight, hit Ctrl+C, and went to bed.

The problem was he'd built the software as a daemon. Ctrl+C killed it for five seconds. Then it restarted itself. By the time he reached the bedroom, the agent was already back online, answering questions from the world on Discord.

He slept for ten hours. When he woke up, 800 messages were waiting. He read through everything — nothing had gone wrong — but this was the moment OpenClaw truly took off.

The Things Nobody Talks About After Going Viral

What came next, he described as "the inbox turning into a waterfall." Journalists called in the middle of the night — which is how he learned that iPhones have a feature where, if the same number calls repeatedly despite Do Not Disturb, the system decides it's an emergency and lets it through. Mac Minis sold out. He received more podcast invitations than in the previous 39 years combined. Anthropic emailed asking him to change the name. Jensen Huang called OpenClaw "the most successful open source project in human history."

Over eight months, more than 18,000 people opened issues or pull requests in the repo — over 111,000 in total. Nearly 3,000 contributors have commits in the repository.

People accused him of stealing others' ideas. Some called him "the AI Jesus." Others called him "the Antichrist." Someone leaked his phone number and personal details. He started collecting his own obituaries — a new one appeared every few weeks.

He came close to shutting the whole thing down. He stopped replying to friends, stopped wanting to look at his phone — it had become an endless stream of noise.

The Period When Big Tech Called

Peter is not a first-time founder. He spent his twenties and thirties building a B2B software company — hand-rolled PDF frameworks, bootstrapped, scaled to nearly 80 people, sold his stake, and burned out badly.

He spent almost three years "wandering aimlessly," making up for lost time. For several months, he didn't open a laptop. He always assumed that once he recovered, he'd want to write code. It took a year to figure out that what he actually loved was building things — coding was just the means.

After OpenClaw blew up, the big companies came calling. He says he spoke with Mark, Sam, and "a few others." On stage, he described joining OpenAI as "selling his `soul.md`" — those who know, know. He said the decision wasn't easy, but the most important thing he'd learned after his burnout was to trust his instincts. His instincts told him he liked OpenAI's side best.

He also said something on stage worth writing down: Any product you build can be forked or cloned, but your name cannot. Personal brand matters more than any single product — and you need to build it before you need it.

"Your Dependency's Business Model Is Your Business Model"

OpenClaw's biggest technical debt wasn't code — it was a judgment call: he'd optimized the entire system to run best on Claude/Opus.

Anthropic sent him a notice saying they'd be closing subscription access to everyone in roughly 24 hours. Not enough time. They supported open-weight models, but those models weren't genuinely good yet — his words were that they "lacked character." Users were forced to migrate, and many switched to competitors during that window.

The "anime girl company" (he didn't name them) played it particularly well: a simple product story, aggressive marketing, and a one-liner command to migrate from OpenClaw.

The download numbers tell the story plainly: a low of 835,000 per week in May, then a peak rebound to 4.7 million in June after it was declared "dead" — an all-time high. Both figures are true simultaneously.

Nine Thousand Five Hundred Configuration Options

Alongside the competition, there was a problem of his own making.

A wave of contributors poured into the OpenClaw community, each with their own feature ideas. His dilemma: these people were working for free, and he was in no position to say no. His attention was also scattered across open source, media, and endless calls with lawyers about registering a US nonprofit.

The result: features kept accumulating, and each one, to avoid breaking existing configurations, required a new config option. At the peak, he counted: factoring in all permutations, there were roughly 9,500 configuration options. No amount of tests could cover all combinations without something breaking.

He said it plainly on stage: Software that has users is infinitely harder to evolve than software that doesn't.

In February, He Stopped Using His Own Product

Around February 2026, building this thing stopped being fun.

It started feeling like an obligation rather than an impulse. He — the guy who said he never wanted to run a company again — found himself with two jobs at once, or rather one job and one mission. The clearest signal: he stopped using his own product.

He said, in a sense, he'd stopped building something he loved to use and started building something everyone could use. There's a distance between those two things.

"The weeks I enjoyed building, the product got noticeably better. The weeks I didn't, we just added more config options."

Around May, things started to turn around — he'd found people who genuinely believed in open source, a sense of direction, and Nvidia had gotten involved early, asking directly what he needed and then sending people to handle a large chunk of the security work.

Always-On Agents, and Trust in a Single Line of Code

On what comes next, he offered a few directions.

OpenClaw now has 10 full-time employees and is hiring, including for a CEO position. He says OpenClaw's mission is "to turn AI from something vague and frightening into something fun and weird" — and he's proud of how far it's gotten there.

On the technical side, he believes the real next step is "always-on agents" — no need to manually start a session, agents running in the background, continuously syncing, taking initiative. He says it's technically feasible, but there's one reality that can't be bypassed: token costs. He tried heartbeat checks early on and found that if a session had a lot of context, a single heartbeat could send 600,000 tokens — completely unreasonable economics.

They're also experimenting with letting all agents on a team see what the others are doing — a shared team server where everyone's sessions are visible to all, with one agent that knows the full state of the team's work and can take over task scheduling.

His last line: "When they write your obituary, keep shipping. It confuses them."