TheChek News Flow

The Takeaway

OpenClaw, the AI agent software that took the tech world by storm, is reckoning with a difficult choice: whether to transition from an experimental hacker project to a sustainable tool companies can reliably use.

One of the fastest-growing open-source software projects ever, OpenClaw is still managed by a group of volunteers who are trying to keep up with the technical and logistical demands of a fast-changing new product category while hewing to the collaborative, experimental spirit that got it started.

Created by Austrian programmer Peter Steinberger, OpenClaw became an overnight sensation by enabling developers and tech enthusiasts to build their own autonomous AI agents. Its popularity, especially in China, prompted tech giants including ByteDance and Tencent Holdings to launch OpenClaw-based products. At Nvidia’s GTC conference last month, CEO Jensen Huang called OpenClaw “the operating system for personal AI,” portraying its impact as akin to the revolution Windows brought to personal computers.

Steinberger and a few dozen volunteer “maintainers” are responsible for adding features, fixing bugs and addressing concerns from its ever-growing list of users. The maintainers typically juggle moonlighting for OpenClaw with day jobs such as running their own startups. They use AI coding tools to work on OpenClaw’s almost constant software updates.

Steinberger got a new day job in February when OpenAI hired him. But the AI giant has given him plenty of time to continue leading OpenClaw’s development, according to two people with knowledge of the matter. In addition to running OpenClaw as an open source project, Steinberger has been fundraising for the OpenClaw foundation he announced in February, which will eventually oversee the software.

OpenClaw’s rapid expansion has exposed security and technical vulnerabilities, with users reporting cases of OpenClaw agents accidentally deleting data, and cybersecurity firms identifying malware targeting OpenClaw. Many users over the past month reported that its frequent software updates—sometimes occurring every day or two—messed up the existing agents they had custom-built with tools from the OpenClaw community.

Earlier this month, OpenClaw’s maintainers found out that the OpenClaw-based products built by Nvidia, ByteDance, Tencent and several other Chinese tech companies were still using old versions of OpenClaw from the first half of March. The companies worried that switching to newer versions would result in serious technical problems, according to the two people and another person with direct knowledge of the situation.

Steinberger and OpenClaw’s maintainers are trying to address these challenges. For example, some maintainers have recently started helping the tech companies using older versions migrate safely to the latest version, according to the three people. While some of those companies began to implement gradual software updates in the past week, the migration will likely take time, the people said.

Those challenges are forcing OpenClaw to confront fundamental questions about how it operates, sparking debates among maintainers, according to two of the three people. Some maintainers advocate adopting a more traditional product-management model with predictable update cycles, rigorous quality assurance and formalized coordination with corporate users. Others fear introducing conventional procedures and rigid structures could stifle innovation and diminish OpenClaw’s hackerlike ethos, which values rapid experimentation and free collaboration.

Last week, OpenClaw had a “no new features” week, focusing instead on improvements to make the software faster and more stable, said Gustavo Madeira Santana, a neuroscience doctoral student at Yale University, who has been volunteering as an OpenClaw maintainer since January. “We want to increase stability and make sure that people don’t get scared about updating,” he said.

Some maintainers say OpenClaw needs to update its software frequently to keep reinventing itself, given the ongoing evolution of personal AI agents and users’ expectations. An updated version from just days ago may already require a complete makeover. In their view, the experimental nature of OpenClaw makes it unsuitable for conventional software rollouts.

Established open-source projects have worked through similar challenges. Linux, the popular operating system launched in the early 1990s, has “long-term support” versions that receive small updates to fix bugs but mostly stay consistent so companies can trust they will remain reliable for years.

Darian Shirazi, a general partner at venture firm Gradient who has used OpenClaw, suggests it could have a revenue-generating “professional edition” that it sells in parallel to the open-source project. That would introduce an incentive to make the product more consumer friendly, an approach other open-source projects have taken, he said.

Conflicting Demands

OpenClaw’s maintainers are also dealing with conflicting demands from users. Earlier this month, a software update caused users to start seeing notifications requesting their permission for agents to perform certain actions. Many users who wanted their agents to work autonomously without intervention complained. In the end, the maintainers worked on another update to hide the notifications.

“To understand OpenClaw, you need to understand that the beginning was irregular,” said Noam Schwartz, co-founder and CEO of Alice, a cybersecurity company, who has been using OpenClaw agents for months. “When OpenClaw went out, it behaved more like a product than an open-source project,” leading to unrealistic expectations, he said.

When Schwartz updates to the latest version of OpenClaw, it usually takes him about an hour to fix the various ways his setup breaks. But that is typical of open-source software, he said. “It’s not like an iOS update.”

Anthropic has also been causing headaches for OpenClaw lately. Anthropic earlier this month announced new rules that made it much more expensive for customers to use its models with third-party tools like OpenClaw. As a result, instead of paying $200 per month for an Anthropic subscription, “you can end up paying $200 per day,” said Schwartz.

The increased cost is a blow to OpenClaw users, many of whom rely on Anthropic’s Claude as their preferred AI. But OpenClaw is working to help people make better use of OpenAI’s models instead, said Madeira Santana.

OpenClaw’s popularity has drawn competitors. Hermes, an agent tool from Nous Research, earlier this month eclipsed OpenClaw in terms of recent contributors, according to GitHub data compiled by ClawCharts, which tracks the growth of open-source AI agents. Hermes is gaining in terms of other metrics, such as new software changes and stars (GitHub’s version of likes)—although Madeira Santana said OpenClaw is still receiving as many new code proposals as ever.

Shirazi of Gradient said he has found OpenClaw to be clunky, and he also worries about its security, so he has switched to rivals such as Town. “I don’t know why anyone that isn’t an engineer and isn’t a tinkerer would be using OpenClaw at this point because there are a lot of these solutions, like Town or Notion custom agents, that are really good,” he said.

The fragmented nature of open-source software makes it difficult to gather comprehensive data, but statistics from Node Package Manager, software for downloading OpenClaw, show average weekly downloads have fallen by roughly half since the peak in mid-March, returning to lower levels it had earlier that month. Those numbers include updates of existing OpenClaw software as well as new downloads, but they don’t count downloads via other platforms.

As they work to build out OpenClaw, its maintainers are getting help from established companies. Tencent recently agreed to contribute to the OpenClaw foundation, according to two of the people with knowledge of the situation. OpenAI offered to donate to the foundation as well (Steinberger has said he turned that offer down), and it helped with the organization’s paperwork and has provided credits for OpenClaw maintainers to use OpenAI’s Codex coding tool, according to an OpenAI spokesperson.

GitHub also has offered complimentary services to OpenClaw maintainers, and Convex, which provides software and databases to host websites, covers the bill for ClawHub, where OpenClaw users download new skills for their agents.