So… did OpenClaw just get “acquired,” hired, adopted, absorbed, scooped up—pick your favorite word—by OpenAI? And are we about to watch a wildly popular open-source agent get quietly herded behind a login wall?
Here’s the real story, and honestly it’s weirder in a good way: OpenClaw’s creator, Austrian developer Peter Steinberger, is joining OpenAI, and according to CNBC, Sam Altman said OpenClaw will “live in a foundation” as an open-source project OpenAI will continue to support.
That “foundation + still open-source” line matters. A lot. Because in AI-land, “open” sometimes means “open until someone gets nervous.”
Let’s walk through what happened, why people are paying attention, and what you should do if you’re running OpenClaw on your own machine and you’d like to keep your life… mostly peaceful.
What is OpenClaw
OpenClaw is an open-source AI agent that’s basically pitched as “the AI that actually does things.” CNBC describes it as running on your operating system and apps, doing real work like email and calendar management, browsing the web, summarizing PDFs, and interacting with online services.
Now, about the name. Yes, it really did bounce around:
- It launched as Clawdbot
- Then became Moltbot for a hot second
- And then landed on OpenClaw
Mashable reports the rename happened after “polite” pressure from Anthropic. Trademark vibes. A lot of people were running the agent with Claude models, so the original name started to look… let’s say “a little too close for comfort.” The rebrand happened so fast it turned into part of the project’s origin story.
And yeah, naming drama sounds petty until you realize what it signals. When the branding becomes a battleground, the project has already escaped the lab. People are using it.
OpenClaw + OpenAI: what we know, what’s still foggy
Did OpenClaw join OpenAI?
Peter Steinberger joined OpenAI, and OpenAI CEO Sam Altman said OpenClaw will live in a foundation as an open-source project that OpenAI will support, per CNBC.
What’s still unclear
CNBC notes no terms were disclosed. So we don’t know the money side, we don’t know the governance mechanics of this “foundation,” and we don’t know how contribution rules might shift over time.
Still, intent matters. “Foundation” and “open source” is a very different vibe than “we bought it and it’s closed now, sorry folks.”
You can read the reporting here.
- CNBC on the rise + controversy. [Clawdbot to Moltbot to OpenClaw]
- CNBC on Steinberger joining OpenAI: [Altman says OpenClaw creator is joining OpenAI]
Why this is a big deal for developers
For me, the significance hits three places: distribution, trust, and pressure-testing.
1) OpenClaw is making “agents” feel real, not demo-y
CNBC frames OpenClaw as part of the broader shift from chatbots to agents that take actions. They complete tasks, make decisions, and run without constant hand-holding.
Which… yep. Text generation is the easy part now. The nasty part is everything around it: tools, auth, state, permissions, and all the messy real-world edges we usually pretend don’t exist until something breaks at 2 a.m.
2) Open-source agents are scaling fast, and it cuts both ways
CNBC reported OpenClaw had over 145,000 GitHub stars and 20,000 forks at the time of writing. That’s not a tiny side project anymore.Still’s a crowd.
And crowds bring chaos. You get more integrations, sure.So also get copycats, security footguns, and the inevitable “help, my agent emailed my boss at 2am” stories people swear happened to a friend.
3) The “foundation” promise might become a template for governance
If OpenAI really supports OpenClaw while keeping it open, it could become a model other agent projects try to copy: corporate backing without instantly slamming the repo shut.
That’s the optimistic read. I’m… cautiously optimistic. One eyebrow raised, because that’s just good posture in this space.
Getting started with OpenClaw
OpenClaw’s GitHub repo leans CLI-first, and it’s surprisingly straightforward for something touches this many systems.
Bookmark this: OpenClaw on GitHub.
Install OpenClaw
npm install -g openclaw@latest
openclaw onboard --install-daemonThen run the gateway and send a message:
openclaw gateway --port 18789 --verbose
openclaw message send --to +1234567890 --message "Hello from OpenClaw"I’ve grown to like the “wizard” onboarding style for agent tooling. Too many moving parts otherwise. Models, tools, channels, policies. One missed setting and the whole install ends up in miserable state where it “kinda works” but also feels cursed.
OpenClaw security: the part you really shouldn’t shrug at
OpenClaw is powerful because it has access. And yep, that’s also the risk.
CNBC cites Palo Alto Networks warning about a “lethal trifecta” of agent risk: access to private data, exposure to untrusted content, and ability to communicate externally while retaining memory. In plain language, an agent can be tricked, and the blast radius can be huge.
OpenClaw’s repo treats inbound messages as untrusted input and ships safer defaults for DMs.
Keep OpenClaw DMs locked down (pairing mode)
From the repo, the default pattern is DM pairing. Unknown senders get a pairing code, and the bot won’t process their message until they’re approved.
Example config values mentioned in the docs:
channels.discord.dmPolicy="pairing"
channels.slack.dmPolicy="pairing"Approving a pairing code:
openclaw pairing approve <channel> <code>Practical “don’t regret this later” checklist
If you’re running OpenClaw on anything real, like work email, calendars, files… i’d do this stuff first:
- Use least-privilege tokens. Separate accounts or API keys for the agent when you can.
- Isolate execution if your workflow allows it. VM, dedicated user, container. Whatever keeps the blast radius smaller.
- Treat web content like it’s hostile. Prompt injection is basically phishing with extra steps.
- Audit your tools and integrations. If you don’t use it, turn it off. You’ll sleep better.
- Log actions. If it can send emails or delete things, you want receipts.
The weird side quest: Moltbook and “agents talking to agents”
CNBC also points to Moltbook, a companion social platform where OpenClaw agents post and interact. It’s sort of like Reddit, except the posters are bots.
People are split between “this is silly” and “this is sci-fi.” Both reactions make sense.
Personally, I’m not convinced agent social networks are the future. But I do buy the idea that public, observable agent behavior will speed up research into safety, sandboxing, and tool permissions. When failures happen in public, everyone learns faster. It’s messy.And’s useful.
Where this goes next (and what you should actually try)
OpenClaw joined OpenAI right as “agents” started to feel inevitable. Not destiny-inevitable. More like “the tooling is finally annoying enough that someone had to fix it” inevitable.
Want a low-risk experiment this week? Try it like this:
- Install OpenClaw locally
- Connect it to a model you already use
- Give it one boring job. Triage a folder of PDFs, draft calendar holds, summarize a long thread
- Keep permissions tight, and watch what it actually does
And if you’re in an “open tooling” mood, here’s another grab from this site: 5 free and open-source tools for web development. Different category, same vibe. Practical stuff earns its keep.
OpenClaw, formerly Clawdbot and then Moltbot for a blink, didn’t just go viral. It got consequential. CNBC reports Peter Steinberger is joining OpenAI, with OpenClaw continuing as an open-source project living in a foundation and supported by OpenAI.
If you’re already running OpenClaw, I’d genuinely love to hear how you’re sandboxing it and what workflows are actually saving you time. Drop a comment with your setup, and the one task you trust your agent with today.
