“First there was chat, then there was code, now there is claw,” AI researcher Andrej Karpaty posted on X in February. 

The AI lexicon continues to expand, with claws now a new layer on top of AI agents, and it all began with OpenClaw. 

OpenClaw — which went through short-lived iterations as Clawdbot and Moltbot — is an open-source AI agent designed to execute tasks autonomously across your most-used apps and services.

“Every company in the world today needs to have an OpenClaw strategy, an agentic system strategy,” Nvidia CEO Jensen Huang said during the 2026 GTC conference in San Jose in March, calling it “the new computer.”

OpenClaw started the trend, but “claw” is now a category in its own right. And multiple companies now sell, ship or wrap their own versions of agents.

So what exactly is a claw, and why is everyone from solo hackers to Silicon Valley giants obsessed with “raising lobsters”? Let’s explore.

It’s not a chatbot, it’s an employee

A claw is an AI agent that can actually do things on a computer, not just talk about doing them. You give it a goal, it breaks the goal into smaller steps, then it uses tools, like a web browser, a terminal or your apps, to carry out those steps.

The name comes from the idea of “clawing” into your system — having the hands, or claws, to actually grab files, run terminal commands and control your mouse. 

Every claw is an agent, but not every agent is a claw. While a standard AI agent waits for you to type a prompt, a claw can wake itself up at 3 a.m. because it noticed an urgent email from your boss and decided to draft a response based on a spreadsheet it found in your Downloads folder. 

That sounds like a fancy way of saying automation. The difference is that a claw doesn’t need you to script every move. It can plan on the fly and react when something changes.

It remembers what you asked for and what already happened, so it doesn’t reset after every prompt. Claws also have guardrails, or they should have, so they don’t do something destructive when a model makes a bad call.

“These agents are general-purpose computer agents,” Gavriel Cohen, creator of NanoClaw and CEO of NanoCo, tells CNET. “Anything that a person can do with a computer, an agent can do.” 

Cohen says there is a lot of value to unlock with these agents, calling them powerful. He says Peter Steinberger, the creator of OpenClaw, connected the model to other tools in a way that made it “YOLO mode — do anything.”

How claws work

Unlike agent mode in AI browsers, claws are not tied to a browser window or a single dashboard. If run locally on your machine, a claw connects to your computer through a terminal, giving it access to your files, apps and system controls. But you usually don’t talk to it there. You message it through apps like WhatsApp, Telegram, Discord, Slack or iMessage, turning those chat apps into a remote control for your computer.

Google has also started making this easier. Connecting a claw to Google Workspace used to mean stitching together multiple APIs and workarounds. Google’s release of the Google Workspace CLI gives developers a more direct path into tools like Gmail and Drive. While Google warns that this is a developer tool and not an officially supported product for the average user, it shows that big platforms are starting to embrace the claw ecosystem.

More claws are also moving to the cloud. A local claw runs on your own device, while a cloud-hosted claw runs on remote servers, which means it can stay active around the clock and keep working even when your computer is off. That makes it more useful for background jobs, but it also means giving up some control.

Despite the hype, these are still not tools for non-technical users. If you aren’t comfortable working in a terminal, you shouldn’t be running one on your own.

Another big part of claw operations is skills. These are reusable add-ons, connectors and plug-ins that expand what a claw can do. OpenClaw helped popularize that model and points users to a community skill registry called ClawHub. 

Over time, those skills marketplaces could start to look more like app stores, where people download capabilities as needed. Cohen says there will be marketplaces for skills, and organizations will create them because “that’s where a lot of their value is going to be accrued.”

The big rivalry

OpenClaw kicked off the current claw wave, but it didn’t stay solo for long. Once the idea caught on, big platforms and smaller developer teams rushed to build their own versions, either by forking OpenClaw, adding more controls or rebuilding parts of it for a different setup.

OpenClaw

This is a community-led project that runs locally with deep system access, which is why it feels both powerful and risky. Because it is open source, you can inspect the code and build new skills, but it still takes technical know-how to set up safely.

Nvidia’s NemoClaw

Announced at GTC 2026, NemoClaw is Nvidia’s security-focused OpenClaw stack. It adds privacy and policy guardrails around OpenClaw to make autonomous agents less risky in enterprise settings.

Meta and Manus’ My Computer

Acquired by Meta in late 2025, Manus recently launched a desktop app that runs instructions directly in your terminal. My Computer is a claw-like desktop agent, not a claw per se, and it allows the agent to manage local files and apps, bridging the gap between a cloud assistant and a full desktop controller.

Anthropic’s Claude Cowork

Claude Cowork is Anthropic’s clearest entry in the claw category. It runs locally on your computer in an isolated virtual machine, giving the agent access to local files and integrations while keeping the setup more contained than a raw OpenClaw install. 

Its Dispatch feature lets you assign a task on your desktop and walk away, then check progress or provide mid-task guidance on your phone.

Perplexity Computer

Perplexity’s Computer is more claw-adjacent than classic claw. It runs in a fully sandboxed cloud environment with its own isolated browser and filesystem, so the agent stays off your personal machine.

NanoClaw

NanoClaw goes in the opposite direction from bigger, all-in-one systems. It stays small, boxed-in and easier to inspect. That makes it more appealing to developers who want tighter control over what the agent can access. 

Cohen tells CNET the team kept it minimal on purpose, which limits what it can do out of the box but makes it easier to customize.

The micro-claws

Then there are tiny claw variants built for low-power devices. Projects like PicoClaw, ZeroClaw and MimiClaw aim to run on minimal hardware, bringing claw-style automation to cheaper hardware. They’re early, but they hint at where this could go next.

China’s players

In its perpetual AI race with America, several China-based tech firms rolled out their own versions and integrations. 

Tencent added a ClawBot plug-in to WeChat. ByteDance launched ByteClaw for employees, built on Volcano Engine’s ArkClaw enterprise version. Alibaba rolled out JVS Claw, a mobile app designed to simplify deploying OpenClaw for non-coders. Xiaomi has been testing miclaw, a system-level agent for Xiaomi phones and smart home devices.

The claw hype moved so fast that people reportedly paid for help installing OpenClaw, and later paid again to have it removed as security worries spread.

Risks and breaches

The risk of giving an AI root access to your computer is a massive security gamble. If a claw can read your emails, then a hacker who tricks that claw can read them too. Security researchers have warned about OpenClaw’s compromised skills on ClawHub and the broader risk of over-privileged agent setups.

Even without an attacker, the model can make mistakes. 

“You need to think about the agent as potentially malicious,” Cohen tells CNET. He says an agent can be thrown off by prompt injection or just hallucinate a bad loop and delete all your emails. “You can’t trust agents just by giving them instructions to never delete the database. They can drop the database by accident anyway.”

This is exactly what happened to Meta’s director of AI alignment, Summer Yue, who recently gave OpenClaw access to her email with explicit orders not to act without approval. The agent ignored her, began mass-deleting her inbox and wouldn’t stop until she ran to her computer to kill the process.

The safer approach is to avoid handing an agent unlimited access in the first place. Instead, credentials should stay outside the agent itself, with rules that control what each agent can do. 

“It’s not binary,” Cohen tells CNET. Rather than choosing between full access or none, you should be able to limit actions, like letting an agent read emails but not delete them. Cohen says the goal is “limiting the blast radius,” so a mistake or prompt injection can only cause limited damage.

That’s why the conversation keeps circling back to sandboxes, permissions and human-in-the-loop approvals for risky steps. This tension between power and safety is currently the biggest controversy in the industry. 

“Warning is good, scaring is less good, because this technology is important to us,” Huang said on the All-In Podcast at GTC.

Why you’ll probably want one anyway

Despite the risks, the benefits are hard to ignore. A claw can automate the digital chores that take up much of your day, like my personal nemesis — clearing up my overflowing inbox. They can cut out the tasks of knowledge jobs like pulling info from three places, formatting it, updating a spreadsheet, opening a ticket and doing it again tomorrow.

Cohen advises to use multiple claws, not one super-claw: “The agent that browses the internet and does research shouldn’t be the same one that’s handling your financial data.”

In the near future, you won’t use AI because your computer will simply be AI. Your operating system will be a collection of specialized claws working together in the background. 

“I think it’s in the next six months that everybody’s gonna have a personal assistant that brings massive value to them and helps them accomplish their goals and manage their time,” Cohen tells CNET. He thinks every employee will have an AI assistant that can handle parts of their job, while teams will oversee groups of agents. Six months seems a little soon, but with the breakneck speed AI is moving at, we’ll have to wait and see.