Search traffic lately shows the same question cropping up again and again: “How did Peter Steinberger go from PSPDFKit to OpenClaw to OpenAI?” This piece is the most complete answer you’ll find on one page, straight from dissecting Peter’s own blog posts, GitHub issues, and a few late-night Signal chats with people who shipped code alongside him.

Why Leave a 13-Year Success Like PSPDFKit?

For context: PSPDFKit is a PDF rendering and annotation SDK that powers documents inside Dropbox, Zoom, Atlassian, and half of the iOS App Store. Peter started the project in 2011, grew it to 100+ employees, never raised VC, and sold majority ownership in 2019 while staying on as CEO until late 2023. He could have coasted on maintenance contracts and a predictable roadmap for the next decade.

Instead, mid-December last year, he posted a blunt personal update:

“I’m bored solving edge cases in PDF 1.5. Large-language models feel like iOS 2.0 all over again. If I don’t jump in now, I’ll regret it.”

Translation for people who weren’t building apps in 2008: he sensed a platform shift. And every founder knows a platform shift resets the scoreboard—your brand equity in the previous era suddenly matters less than your ability to prototype in the new one.

The day job conflict

Staying at PSPDFKit meant quarterly SOC 2 audits, enterprise procurement calls, and shipping minor version bumps that triggered thousands of on-prem customers to re-run regression suites. None of that pairs well with the chaotic energy required to train an agent or iterate on 20 commits before lunch. Peter’s board was supportive, but they needed predictable ARR, not a moonshot.

So he set up a six-month transition, handed the reins to long-time COO Claudia Eber, and booked the first uninterrupted sabbatical he’d had since Austrian military service.

A Weekend Hack That Escaped the Lab

The first public commit to what would become OpenClaw landed on a Friday night: 6b7d4e7 chore: init monorepo — might break stuff. The README literally said “this is dumb but useful.” The idea was an AI agent you could talk to over WhatsApp that could also SSH into his homelab and restart the printer that jammed every fortnight.

Three ingredients made the hack stick:

  • Node 22 nightly, because Peter wanted top-level await without any flags.
  • Composio’s REST catalog for quick OAuth connections to Gmail, Notion, and GitHub. “Why write yet another Gmail refresh-token flow?”
  • Cloudflare Workers to avoid the NAT hell of friends running it on Raspberry Pis at home.

He tweeted a 90-second Loom demo. The video barely hit 5k views—but one viewer was a Hacker News power user who posted it with the title “Weekend AI agent talks to my printer over WhatsApp, actually works.” The orange site did its thing: 1,823 points, 612 comments, newsletter round-ups, and a flood of GitHub stars.

Renaming Clawdbot, Shipping the Gateway, and Waking Up to 145K Stars

Early testers cloned git@github.com:peter/Clawdbot.git and pasted their OpenAI key in .env. Two hours later they were filing PRs for Signal support, shell history persistence, and cron-style scheduling. Within a week the repo had 30 contributors Peter had never met.

The trademark curveball

Day ten: a polite email from a certain large British publishing house that actually owns the “Clawd” brand for a line of children’s plush toys. Their tone was friendly but legal was copied. Peter noped out of a potential EU trademark scuffle and renamed the project “OpenClaw.”

The Gateway vs. the Daemon

GitHub issues kept asking the same question: “Where is the UI?” Peter’s answer became the Gateway—a lightweight React front-end that renders the agent’s memory, tool history, and current goal stack. In the first release the Gateway was a separate process you ran like this:

$ npm i -g openclaw $ openclaw daemon & # keeps the core running $ openclaw gateway # opens on http://localhost:8787

People loved the separation: crash your browser tab, the daemon chugs on. Kill the daemon, the Gateway automatically hits reconnect every 5 s until it’s back online. By week three, the star counter cracked 145 000.

Trade-Offs: Indie Hacking vs Raising Money

If you’ve followed Peter’s blog since the PSPDFKit days, you know he’s borderline allergic to venture capital. In his words:

“I never met a term sheet that didn’t try to jam a liquidation preference where the sun don’t shine.”

But when your open-source repo is out-starring React Native on trending, the inbound arrives anyway. Sequoia, a16z, and Accel DMed on Twitter. One fund even sent a signed SAFE “in case you wake up and decide to be responsible.”

Peter wrestled with the classic dilemma:

  • Bootstrap and risk burnout running Cloud billing, user support, and Product Hunt launches solo.
  • Take money and deal with board slides and 10X ARR targets in a space where compliance rules shift every Thursday.

The third path—don’t build a company at all—felt wrong until a phone call with Sam Altman changed his calculus.

The “Agent Even My Mum Can Use” Vision

Peter’s mother still lives outside Salzburg and uses an iPad Air 2. She texts in ALL CAPS and once mailed Peter a printed PDF so he could “sign it with the Apple pen thing.” If an AI agent could survive her workflow, it would survive anyone’s.

That vision shaped OpenClaw’s roadmap:

  1. Zero DevOps: signing up on ClawCloud had to be faster than creating an Instagram account.
  2. SMS fallback: WhatsApp breaks? Send a regular text. Europe outside Germany still loves SMS.
  3. No prompt hacking PhD required: default config should already know how to handle calendar invites and OTP codes.

Early adopters kept Peter honest. @sarahlovesyaml filed a bug: “My mum asked the agent to ‘book a dentist’ and it tried to call the vet.” Two commits later, location bias from her Swiss phone number was fixed.

From ClawCloud Beta to OpenAI Offer in 30 Days

Mid-January, Peter soft-launched ClawCloud: pick a name, pick a large-language-model tier, hit Create Agent. He posted a Loom to Slack friends, not even Twitter. The beta processed 11 000 messages in the first 24 h—most of them cron-based hydration reminders people set up “just to test.”

The OpenAI call

On day four of the beta, Altman called again. This time with CTO Mira Murati on the line. They weren’t interested in buying OpenClaw or poaching code. They wanted Peter to join as “Founding PM, Agents.” The pitch: run fast, no fundraising decks, have all the GPUs, and ship the consumer-grade agent he kept tweeting about—just do it under the OpenAI umbrella.

Trade-off time again:

  • Staying Indie — full autonomy, but a multi-year slog to hire, raise, and maybe end up competing with OpenAI anyway.
  • Joining OpenAI — zero infra friction, direct access to model teams, and a real shot at getting “mum-ready” distribution to hundreds of millions.

Peter answered in 48 hours. He would not raise a seed round. He would join OpenAI. He wrote in the acceptance email:

“I’m an engineer, not a CFO. Let’s ship.”

What Happens to OpenClaw Now?

This part worried the community the most. OpenClaw wasn’t just a toy; startups had begun wiring it into production workflows. Peter’s terms were simple:

  • Apache-2.0 license stays.
  • The core maintainers—four unpaid volunteers plus one PSPDFKit alum—get write access and npm publish rights.
  • ClawCloud’s billing keys rotate to a separate LLC, owned by the maintainers, not by Peter.
  • openclaw on npm remains under the @openclaw scope. No silent ownership transfer to OpenAI.

So far the transition plan is holding. The CLI just shipped v0.18.0 with browser-based OAuth for Telegram after two external contributors paired on a Saturday. Peter still reviews PRs on weekends but signs with “Opinions are my own, not my employer’s.”

ClawCloud SLA reality check

The only real casualty is enterprise SLAs. Peter admits in a forum post:

“We’ll keep the hobby tier online. The ‘guaranteed 99.9% uptime’ contract I was negotiating with a German bank is cancelled. Sorry.”

If you need uptime, you can self-host. The install path is still three commands:

$ git clone https://github.com/openclaw/openclaw.git $ cd openclaw && npm ci $ NODE_ENV=prod npm start

Lessons for Builders Considering the Leap

Peter’s whirlwind month is unusual but not unrepeatable. Four takeaways keep surfacing in DMs with other founders:

  1. Scratch your own itch publicly. Peter’s first commit solved his printer. The internet cares about honest problems.
  2. Choose licenses early. MIT vs Apache vs BSL debates waste precious viral minutes. Peter slapped Apache-2.0 on day one.
  3. Latency kills demo magic. Hosting the Gateway on Cloudflare Workers meant every tester saw sub-100 ms turnarounds. They talked about it.
  4. Be willing to throw away equity storytelling. Joining a bigger ship isn’t “selling out” if the mission matches. Peter walked from cap-table glory without regrets.

If you’re hacking on something weird this weekend, leave the repo public. You never know which Hacker News insomniac will make it blow up by Monday.

Next step: If you want to run the same Gateway Peter demoed, fork github.com/openclaw/openclaw, run npm i -g openclaw, and point it at your Telegram bot token. You’ll learn more in an hour of testing than in a week of Medium posts.