Live in the future, build what's missing.: Who the hell is Peter Steinberger?Artificially generated
Artificially generatedArtificially generatedArticle 50 of the AI Act requires artificially generated image, audio and video content to be marked as such, detectably and machine-readably. The symbol is the European Commission's official base icon. Using it is voluntary, the marking duty is not.Machine-generated: Text Β· Image Β· Video. With human direction.McGrinsey Living Intelligence SystemsMcGrinsey Living Intelligence SystemsLI: Living intelligence of every kind. McGrinsey combines conventional intelligence and novel intelligence into living intelligence systems. These symbiotic systems of machine and human intelligence work together to supply the finest distillates of insight, for every intelligence.
🦞 WHO THE HELL IS PETER STEINBERGER 🦞 CLAW CLAW CLAW 🦞 LET THE MODEL COOK 🦞 FUN IS VELOCITY 🦞 OPENCLAW OPENCLAW OPENCLAW 🦞 WHO THE HELL IS PETER STEINBERGER 🦞 CLAW CLAW CLAW 🦞 LET THE MODEL COOK 🦞 FUN IS VELOCITY 🦞 OPENCLAW OPENCLAW OPENCLAW 🦞
🦞 CLAW
APPROVED
PARTY
MCG Research · Who the Hell is Peter Steinberger · Episode 02 · August 2026 🦞

Live in the future, build what's missing.

A man builds a WhatsApp hack in one hour because he's hungry and annoyed. Eight months later, Nvidia CEO Jensen Huang calls his project the most successful open-source project in human history. We checked what's true about that, and what lies beneath the surface of his own narrative.

From PDF framework bootstrapper to the man whose software allegedly co-caused a Mac Mini bottleneck. The closing sentence of his talk is the key to the entire performance: live slightly ahead of the present, and build what's missing. We show where this attitude actually explains his decisions, where his own narrative is supported by independent journalism, and where it isn't.

ζ˜ εƒ · The Video · 🦞 the entire talk

Everything in this report is based on this appearance. Watch it before or after you read, then you'll notice yourself what we've contextualized below and what remains his own, unchallenged version.

01 · Before he built Claws 🦞

An Austrian PDF nerd who once burned himself out

Before anyone spoke of "Claw," Peter Steinberger was the founder of PSPDFKit, a PDF software company he bootstrapped while waiting for a US visa. The company grew to around 80 employees, and according to their own statements, their technology eventually ran on a billion devices.

In October 2021, investor Insight Partners entered with growth financing of over 100 million euros. This is independently documented, including directly by Insight Partners themselves. Steinberger and co-founder Martin SchΓΌrrer subsequently stepped back from day-to-day operations. In his own narrative, this sounds like a classic exit; the press release reads more like a growth round with withdrawal from operational roles. Both are not mutually exclusive, but it's worth knowing the difference.

What came after, he openly tells himself: around three years in which he barely programmed, instead catching up on life he had foregone while building the company. This burnout phase is consistently recounted in many interviews, but we could find no independent documentation for it, only the same, often repeated self-disclosure.

Real and publicly verifiable, however, are his accounts: GitHub under steipete, his own site steipete.me, X under @steipete. No anonymous Discord myth, a man with a real name and history.

≈80
Employees at PSPDFKit
Before the Insight Partners entry in 2021, bootstrapped without initial VC money.
100+M €
Insight Partners investment
October 2021, independently documented, not a sale in the classic sense.
≈3years
Self-reported break
Consistently repeated, but only documented by himself.
1hr.
Until the first WhatsApp hack
That long, he says, his agent needed for the first prototype of what would later move the world.
02 · The Origin Story 🦞🍳

Annoyed, hungry, a kitchen raid, and an agent that took an hour

His own explanation for the origin is disarmingly banal: while a coding agent was working for him, he wanted to go to the kitchen briefly but had no way to check from his phone how far along the agent was. So he had the agent build the solution itself, a simple relay between his computer and WhatsApp. Done in about an hour.

What's remarkable is the structure of this story, not just the content: he explicitly names being annoyed twice as his source of inspiration, once at the very beginning for the very first hack, and once at the very end of the talk when he describes what still drives him today. Eight months of world history lie in between, but the mechanism remains exactly the same.

According to his own account, the WhatsApp relay became a project with its own name over weeks. When someone unsolicited sent a pull request with Discord support, it became more than a feature, it became the moment he himself describes as a turning point: from a one-person tool it became a community project.

Important for context: the core idea of controlling an AI agent via chat instead of terminal was by no means new at that time; many were building similar things. What apparently distinguished his version was less the technology than the feeling: proactive, reduced to the essential, deliberately calibrated outside standard response patterns so it feels less like a tool and more like a counterpart.

03 · The Name Battle πŸ¦žβš”οΈ

Three names in four days, and we're not sure where the count begins

This is where it gets fiddly, and we'd rather say that openly than smooth it over. Multiple independent sources (including Forbes) agree that the project was originally called Clawdbot, named after "Clawd," an unofficial mascot for Claude Code. The official Wikipedia entry for the project tells a different starting story, with an earlier name "Warelay" and a precursor called "Clawd." Both versions cannot simultaneously be the whole truth, and we could not conclusively resolve this before publication.

Early phase
Warelay / Clawdbot
Inconsistently documented. A named project emerged from the WhatsApp relay hack, but which name came first is disputed between sources.
January 27, 2026
Moltbot
Anthropic asserts a trademark claim on the Claude reference in the name. Renaming occurs on short notice, well documented including by Forbes.
January 30, 2026
OpenClaw
Three days later the final name, because "Moltbot" is hard to pronounce. Stable since then, own Wikipedia entry, own GitHub organization.
What's remarkable: Anthropic is portrayed almost affectionately in his own narrative, not a lawyer's letter but an email he himself describes as "really nice." This stands in interesting contrast to another point further below, the sudden end of discounted API access. The same company appears in his story once as a friendly lawyer and once as the trigger of a serious business crisis.
04 · The Night It Went Viral πŸ¦žπŸŒ™

A system service was to blame, not a marketing plan

This is perhaps the most beautiful punchline of the whole story, and it's in a technical detail. He had set up his project as a launchd service, configured so it automatically restarts after a crash. In the first week of January 2026, he let people interact with his agent in a freshly created Discord server, late into the night.

When he finally wanted to quit with Ctrl-C, the service did exactly what it was built for: it restarted itself after a few seconds and continued answering messages while the builder slept. Around 800 messages accumulated that night, with no one at the wheel.

No launch event, no announcement, no PR plan. A reboot mechanism that did exactly what it was built for, at the most inopportune time for control and the most opportune time for virality.

What came after is independently verifiable: a real Mac Mini shortage, reported among others by Decrypt, though with the fair note that the cause is disputed and should not be monocausally attributed to a single project. And a quote from Nvidia CEO Jensen Huang on a CNBC appearance in March 2026, who publicly used superlatives, verbatim documented and independently verifiable.

🦞 800 messages overnight
🦞 Mac Mini sold out
🦞 Jensen Huang on CNBC
05 · The Dependency Lesson πŸ¦žπŸ”Œ

An email with 24 hours' notice, and an entire business model falters

His project long ran best optimized with a particular provider model; other models were supported but according to his own account were significantly inferior in quality. When this provider announced with about 24 hours' notice that it would shut off discounted access for everyone, there was barely time for an orderly course correction.

This episode is perhaps the most valuable lesson of the entire talk, far beyond his own project history: whoever ties their product to a single external provider doesn't have their own business model, they're just managing the whims of someone else's. Open weights, he says himself, are significantly better today than at the beginning of this story, but the shock sat deep enough to land as its own agenda item in a conference talk.

06 · To Sell or Not? The Real Twist πŸ¦žπŸ’°

He says he remained independent. The headlines say something else.

This is the most interesting discrepancy we found in this research, and it's subtle enough that it doesn't appear in any of the colorful retellings online. In his own talk, he describes in detail how he was courted by "the big labs," spoke on the phone with people he only hints at by first name, compared the whole thing to selling his own soul, and according to his own account ultimately decided against it. His conclusion: he trusted his gut feeling, and it told him he preferred open.

Independently of this, TechCrunch, CNBC, and Bloomberg reported on February 15, 2026, in agreement, that Peter Steinberger was joining OpenAI. This is not vague speculation; these are three serious newsrooms with nearly identical headlines on the same day. Only a good five months later, in July 2026, did he found the OpenClaw Foundation together with Dave Morin as an American non-profit organization, with OpenAI, Nvidia, and Microsoft as documented supporters.

Both can be true simultaneously; an employment relationship with OpenAI and a parallel independently organized foundation project are not necessarily mutually exclusive. But his own talk mentions this OpenAI step nowhere, even though he speaks extensively about precisely this decision. This is not a lie, but it is a conspicuous gap in an otherwise very open narrative.

Fun is Velocity. That's exactly the one sentence from his talk that we quote verbatim here, because it captures the core of his self-narrative better than any paraphrase: joy in building beats any roadmap.
07 · What is proven, and what is merely claimed πŸ¦žπŸ“‹

The Honesty Table

ClaimStatus
PSPDFKit founding, Insight Partners investment 2021Independently verified
Renaming to Moltbot on Jan 27, OpenClaw on Jan 30, 2026Independently verified, original name before disputed
Jensen Huang calls it the most successful open-source projectIndependently verified, CNBC, March 2026
Mac Mini shortage due to the projectReal, cause disputed among media
18,000+ contributors, 111,000+ issues/PRs, 3,000 committersNot independently verifiable, only his own dashboard numbers
835,000 to 4.7 million weekly downloadsNot independently verifiable
OpenClaw Foundation, 501(c)(3), July 2026Independently verified
"10 people on payroll", open CEO positionNot independently verifiable
Move to OpenAI in February 2026Independently verified (TechCrunch, CNBC, Bloomberg), not mentioned in talk itself
音ζ₯½ · The Track 🦞🎡

A tribute, because this madness deserved music

John built a dedicated track for this man and this madness. Runs on the same cyber TV as the McGrinsey homepage usually does, only this time completely repainted in lobster-Ferrari-orange. Click play.

🦞 In honor of a man who built perhaps the wildest open-source rollercoaster of the year out of hunger and annoyance. Turn it up.

Track preview, click to play
McGrinsey πŸ¦ž TV
08 · Fun equals speed: the operating philosophy underneath 🦞🧠

Yes, the substance is fun, just not meant as mood, but as velocity

"Have fun" and "trust your gut" sound like motivational posters, and that's exactly why it's easy to miss what he actually means: for him, fun isn't a feeling, it's an engineering variable that translates directly into speed. Weeks with genuine curiosity visibly advanced his project, weeks of pure obligation only produced configuration options. Fun equals velocity, that's not a platitude, that's his actual thesis. We searched the talk at exactly this level and distilled fourteen patterns, each in our own words.

Pattern 1
The interface disappears
The technology behind it wasn't radically new, prompt in, answer out. What worked was that he completely hid the operation: no model selection, no session start, no context size to keep in mind. The boundary between terminal and counterpart dissolves, that was the actual breakthrough.
Pattern 2
Emotion as product-market-fit test
He didn't find product-market fit through metrics, but through faces: excitement, fear, even anger from friends he told it wasn't ready for them yet. A strong emotion despite rough edges is often a more reliable signal with early technology than any survey.
Pattern 3
The mistake of building for everyone
The break came when he stopped building a product he himself loves and started building one for everyone. After that, he no longer used his own tool. Your own daily use isn't a nice side effect, it's the sensor that shows what's nonsense and what's missing. Without it, you navigate blind.
Pattern 4
Fun as engineering variable, not as mood
His own finding: in weeks that gave him joy, the product became visibly better, in weeks of pure obligation only configuration options resulted. Curiosity expands the search space, sense of duty narrows it to risk management. Fun is thus not a feel-good factor, but a search algorithm.
Pattern 5
Annoyance as information advantage
The recurring pattern, annoyance leads to a build that becomes something unexpected, is more than "scratch your own itch". Annoyance delivers high-resolution knowledge: exactly when it hurts, what's already been tried, what would feel like a real solution. AI lowers the cost of immediately making a prototype from this annoyance to almost zero.
Pattern 6
Delegation instead of prompt formulation
His current way of working isn't "write better prompts", but to delegate more independent preliminary work to the agent: not a pull request for review, but fully tested results documented with screenshots. The human shifts from operator to judge, evaluating results instead of executing steps.
Pattern 7
Attention is the scarce resource, not code
When anyone can build a prototype, building is no longer the bottleneck. He himself calls attention the most expensive currency of all. Your own enthusiasm thus becomes a real acquisition channel, because it attracts others before any marketing budget kicks in.
Pattern 8
Complexity as the price of AI speed
A new feature is just one prompt away, the actual costs come afterward. At its peak, his project counted around 9,500 configuration options. The question shifts from "can we build this" to "does the system remain manageable after five hundred built things". This is likely to become one of the central problems of agentic software development.
Pattern 9
Open source as leadership trap
You're reluctant to contradict someone who contributes for free. That's exactly where unplanned growth comes from: pull request, merge, feature, configuration, maintenance burden. His answer was a written vision file, important not as a document per se, but as an admission that a maintainer needs an explicit theory of what the project may become, otherwise the community unintentionally becomes the actual roadmap.
Pattern 10
Code review as risk management, not as understanding
He doesn't read every line. Instead, he asks how large a change is, how long it should have taken, whether it belongs to a sensitive subsystem. A sense of how long proper work roughly takes becomes more important than any single self-written line.
Pattern 11
The engineer as manager of agents
His example with a dozen simultaneously running sub-agents that understand a project, break it down, implement, test against each other and inform each other, is no longer autocomplete. It's building a small organization of artificial workers, the unit of work shifts from writing to designing the system that produces code.
Pattern 12
The bottleneck is infrastructure, not intelligence
Permanently running, thinking-along agents don't fail for him due to model quality, but due to power-outlet problems: persistent context costs tokens, caches expire, machines compete for compute time, Mac-specific work can hardly be provided externally. The missing layer is infrastructure for permanent, stateful work on changing machines, not another chatbot.
Pattern 13
Sessions become topics, not containers
Previously, a session was a disposable context window. For him, it's now more like a permanent workspace, clearing it can even be harmful because valuable context is lost. Thought through consistently, "starting a session" should eventually no longer be a concept that users consciously perceive.
Pattern 14
Graphs are repackaged automation
His assessment of workflow graphs is unromantic: input, trigger, decision, action, engineers have always built that. What's new is only that the individual steps can now themselves understand and decide on fuzzy inputs. Celebrating the abstraction itself, he considers the wrong focus.
The common thread through all fourteen patterns: AI shifts software development away from "tell the computer exactly what to do" toward "build systems that figure out themselves what to do". This also shifts where the advantage lies, away from pure programming ability, toward taste, judgment, orchestration and attention. His own path, from hungry annoyance over missing WhatsApp access to non-profit foundation, is exactly this shift in fast-forward.
09 · What remains 🦞

Four lessons, regardless of whether you believe every number

Lesson 1
Annoyance is a product radar
His biggest hack began as frustration over a missing feature, not as market analysis.
Lesson 2
Your name cannot be forked
Every product can be copied, your own person and reputation cannot.
Lesson 3
Your dependency is your business model
A 24-hour email from a single provider can overturn an entire strategy.
Lesson 4
Verify press figures, including your own
Anyone accusing media of exaggeration should be able to substantiate their own counter-number. That's exactly what's missing here.
ε‡Ίε…Έ · Sources 🦞

Everything this report is based on

We deliberately quote the talk transcript itself only minimally and predominantly paraphrase it; it is the foundation, but not the source for the facts below. Every number we could not independently confirm is explicitly stated in the text as unconfirmed, not as fact.

  1. Insight Partners: Press release on PSPDFKit investment
    Primary source for the over 100 million euro growth financing in October 2021 and the founders' withdrawal from day-to-day operations.
    insightpartners.com/ideas/pspdfkit-announces-e100-million-strategic-investment-from-insight-partners-to-fuel-growth
  2. PR Newswire: same investment announcement, second independent source
    Confirms the figures from the Insight Partners announcement through a second channel.
    prnewswire.com/news-releases/pspdfkit-announces-100-million-strategic-investment-from-insight-partners-to-fuel-growth-301389571.html
  3. Forbes (Ron Schmelzer): Renaming from Clawdbot to Moltbot
    Basis for the name timeline in section 03, including date and Anthropic's trademark claim.
    forbes.com/sites/ronschmelzer/2026/01/27/viral-ai-sidekick-clawdbot-changes-name-to-moltbot-and-sheds-its-old-skin
  4. Wikipedia: OpenClaw (project article)
    Divergent origin story with the name "Warelay", used to expose the inconsistency in section 03, not as sole truth.
    en.wikipedia.org/wiki/OpenClaw
  5. CNBC: Jensen Huang calls OpenClaw the most successful open-source project
    Verbatim documented statement from March 2026, basis for section 04.
    cnbc.com/2026/03/17/nvidia-ceo-jensen-huang-says-openclaw-is-definitely-the-next-chatgpt
  6. Decrypt: OpenClaw and the Mac Mini shortage
    Reports the Mac Mini sellout, with fair note on disputed causality.
    decrypt.co/366389/openclaw-apple-mac-mini-shortage-ai-2026
  7. The Hacker News: Koi Security investigation into malicious extensions
    341 of 2,857 reviewed extensions classified as malicious, around 12 percent, basis for the security table in section 05.
    thehackernews.com/2026/02/researchers-find-341-malicious-clawhub.html
  8. TechCrunch: "OpenClaw creator Peter Steinberger joins OpenAI"
    Basis for the inconsistency in section 07, dated February 15, 2026.
    techcrunch.com/2026/02/15/openclaw-creator-peter-steinberger-joins-openai
  9. CNBC and Bloomberg: same OpenAI announcement, two additional newsrooms
    Independent second and third confirmation of the OpenAI move on the same day.
    cnbc.com/2026/02/15/openclaw-creator-peter-steinberger-joining-openai-altman-says
  10. InfoWorld: Founding of the OpenClaw Foundation
    Confirms the non-profit founding in July 2026 with Dave Morin and the named supporters.
    infoworld.com/article/4196376/openclaw-becomes-a-nonprofit-foundation-as-it-seeks-to-be-the-switzerland-of-ai-2
  11. Peter Steinberger: own accounts
    GitHub, personal site and X profile, for classification as a real identifiable person with history.
    steipete.me

McGrinsey's own assessment: Numbers that come exclusively from the talk itself, such as contributor, download and security figures, we have NOT adopted as fact, but clearly marked as unconfirmed self-reported data, even if they sound plausible. The phrase "the anime-like competitor company" from the talk could not be attributed to any specific company and we therefore deliberately leave it unnamed.

Connection 🦞
From Eugene Schwartz to Peter Steinberger: two men, a century apart, the same stubbornness.

This issue is episode 2 of our "Who the hell is" series, after Eugene Schwartz. Both times it was about the same pattern: a reputation that is bigger than what can be cleanly documented, and the work of exposing precisely this gap rather than smoothing it over.

And yes, this issue looks different from the rest of the magazine. That's intentional, not arbitrary, and remains an exception. 🦞