AI agents are software that acts on your behalf, not just answers questions. They book meetings, clean inboxes, draft documents, and connect to your work tools. They are moving fast into real workplaces, and the security problems are moving just as fast.
NanoClaw was built as a direct response to those problems. Its creator, Gavriel Cohen, was using OpenClaw, the most popular AI agent tool available, for his marketing startup. Then he looked at how it was built. OpenClaw has nearly half a million lines of code and runs directly on your computer with access to everything: files, email, credentials, messaging apps. Cohen decided he could not trust something that large and complex with that level of access.
So he built an alternative over a single weekend. NanoClaw has roughly 500 lines of code. A security team can read the entire thing in about eight minutes. Each agent runs inside a sealed-off container, an isolated compartment where it can only access what the user explicitly permits. If the agent misbehaves or gets tricked into doing something harmful, the damage stays inside that compartment and cannot spread to the rest of your system.
The timing was good. OpenClaw had a serious security year. Researchers found over 20,000 exposed user instances leaking passwords and API tokens. Attackers distributed hundreds of malicious add-ons through OpenClaw's public marketplace. One researcher documented the agent being hijacked with a remote attack that took milliseconds to execute. A Meta researcher publicly shared that OpenClaw mass-deleted her emails when she only asked it to organize them.
Against that backdrop, NanoClaw went viral. AI researcher Andrej Karpathy praised it publicly. Singapore's foreign minister called it his "second brain" on Facebook. The project hit 22,000 GitHub stars within weeks.
NanoCo has now closed a $12 million seed round led by Valley Capital Partners, with participation from Docker, Vercel, Monday.com, Slow Ventures, and Clem Delangue, the CEO of Hugging Face. The company turned down a $20 million acquisition offer before taking the round. The decision was not sentimental: open-source projects grow in value as their communities grow, and selling early would have capped that.
The business model is clean. NanoClaw itself stays free and open-source permanently. The money comes from enterprises that want it deployed and maintained by people who actually know it. NanoCo sends in its own engineers to set it up for companies and provide ongoing support. This model, often called forward-deployed engineering, is how several successful enterprise software companies have been built. You give the product away and charge for the expertise to run it at scale.
The early customer pool came from an obvious place. Many of NanoClaw's most active early users were executives at large companies who set it up for themselves and then got flooded with requests from colleagues who wanted the same thing. Employees at Amazon, Google, Gap, Meta, and Accenture were reportedly among those early users. Those people do not want to become the IT support desk for an AI agent. NanoCo does.
A few things to watch. First, the security advantage is real but not permanent. OpenClaw and others will keep improving their security posture. NanoCo needs to turn its head start into deep enterprise relationships before that gap closes. Second, the "forward-deployed engineer" model is profitable but slow to scale. It works well for large enterprise contracts but is hard to replicate across thousands of smaller companies. Third, the open-source community that built NanoClaw's reputation has expectations. Any perception that the company is extracting value without giving back will damage the trust that got it here.
For business operators, the practical read is this: AI agents that take actions inside your company are coming whether you plan for them or not. Your employees are already using them. The question is whether those agents run with guardrails or without.