Claws Explained: From AI Generation to AI Implementation

“First there was chat, then there was code, and now there is claw,” the AI ​​researcher said Posted by Andrei Karpaty on X In February.

the Artificial Intelligence Dictionary It continues to expand, with the Claws now a new layer above the AI ​​agents, and it all started OpenClaw.

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

“Every company in the world today needs to have an OpenClaw strategy, which is a proxy system strategy,” Nvidia CEO Jensen Huang said during the conference. GTC 2026 Conference In San Jose in March, they called it “the new computer.”

OpenClaw started the trend, but the “claw” is now a category unto itself. Many companies now sell, ship, or package their own versions of… Agents.

So what exactly is a claw, and why has everyone, from lone hackers to Silicon Valley giants, become obsessed with “lobster farming”? Let’s explore.

It’s not a chatbot, it’s an employee

The claw is I have an agent She can do things on the computer, not just talk about doing them. You give it a goal, it breaks the goal down into smaller steps, and then uses tools, such as your web browser, terminal, or your apps, to execute those steps.

The name comes from the idea of ​​”sneaking” into your system – having hands or claws to physically pick up files, run terminal commands, and control your mouse.

Every claw is an actor, but not every actor is a claw. While a standard AI agent waits for you to write a claim, the claw could wake itself up at 3 a.m. because it noticed an urgent email from your boss and decided to formulate a response based on a spreadsheet in your downloads folder.

That sounds like a fancy way of saying automation. The difference is that Claw doesn’t need you to write down every step. Can plan quickly and react when something changes.

It remembers what you asked and what actually happened, so it doesn’t reset after every prompt. The claws also have guardrails, or should have, so they don’t do something devastating when the model makes a bad decision.

“These clients are general purpose computer clients.” Gabriel Cohencreative NanoClaw And NanoCo’s CEO tells CNET. “Anything a person can do with a computer, an agent can do.”

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

How do claws work?

Unlike proxy mode in Artificial intelligence browsersClaws is not tied to a single browser window or control panel. If it runs locally on your device, the claw connects to your computer through a terminal, giving it access to your files, applications, and system controls. But you usually don’t talk to him there. You can send a message via apps like WhatsApp, cableOr Discord, Slack, or iMessage, to turn these chat apps into a remote control for your computer.

Google has also started making this easier. Connecting Claw to Google Workspace used to mean integrating multiple APIs and workarounds together. Google’s version of the Google Workspace CLI It gives developers a more direct path to tools like gmail And driving. While Google cautions 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 tentacles are also moving to the cloud. A local claw runs on your own machine, while a cloud-hosted claw runs on remote servers, meaning it can stay active around the clock and continue working even when your computer is shut down. This makes it more useful for back-end jobs, but it also means giving up some control.

Despite the hype, these tools are still not available to non-technical users. If you are not comfortable working in a terminal, you should not run it yourself.

Another big part of Claw Operations is skills. These are reusable add-ons, connectors, and plug-ins that expand what the Claw can do. Help OpenClaw popularize this model and direct users to a named community skill registry ClawHub.

Over time, these skill marketplaces may begin to look more like app stores, with people downloading capabilities as needed. Cohen says there will be markets for skills, and organizations will create them because “that’s where a lot of their value will accumulate.”

Big rivalry

OpenClaw released the current claw wave, but he didn’t stay alone for long. Once the idea caught on, larger 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 and locally run project with deep access to the system, which is why it appears powerful and dangerous. Because it is Open sourceYou can examine code and build new skills, but it still requires technical knowledge to set up safely.

Nvidia NemoClaw

Announced at GTC 2026, NemoClaw is Nvidia’s security-focused OpenClaw suite. He adds Privacy and policy barriers About OpenClaw to make independent agents less risky in enterprise settings.

Meta and manos my computer

acquired before dead In late 2025, Manus recently launched a desktop app that runs instructions directly in your terminal. My computer It is a claw-like desktop agent, rather than a claw itself, and allows the agent to manage local files and applications, bridging the gap between a cloud assistant and a full desktop console.

Anthropist Claude Cowork

Claude Cowork It is the clearest anthropic 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 compact than the initial OpenClaw installation.

The Submit feature lets you set a task on your desktop and go, then check progress or provide instructions mid-task on your phone.

Puzzled computer

Computer confusion It is more adjacent to the claw than the classic claw. It runs in a fully protected cloud environment with its own isolated browser and file system, so the agent stays away from your personal device.

NanoClaw

NanoClaw It goes in the opposite direction from larger, integrated systems. It stays small, contained, and easy to inspect. This makes it more attractive to developers who want tighter control over what a proxy can access.

Cohen tells CNET that the team intentionally kept it minimal, which limits what it can do out of the box but makes customization easier.

Microclaws

Then there are small claw types designed for low-power devices. Projects like PicoClaw, ZeroClaw, and MimiClaw aim to work with minimal hardware, bringing claw-style automation to cheaper devices. They’re early, but they hint at where this could go next.

China players

In its permanent artificial intelligence race with America, there are several… China-based tech companies have rolled out their own versions and integrations.

Tencent has added a ClawBot plugin to WeChat. ByteDance has launched the ByteClaw app for employees, which is built on the ArkClaw enterprise version of the Volcano Engine. Alibaba has launched JVS Claw, a mobile app designed to simplify OpenClaw deployment for non-programmers. Xiaomi Tested miclaw, a system-level proxy for Xiaomi phones and smart home devices.

The claw noise moved very quickly People reportedly paid to help install OpenClawthen later pushed again to remove it as security concerns spread.

Risks and violations

The risk of granting root AI access to your computer is a huge security gamble. If the claw is able to read your emails, the hacker who tricks the claw can read them too. Security researchers have warned OpenClaw Skills at Risk on ClawHub and the broader risk of overprivileged proxy settings.

Even without an attacker, the model can make errors.

“You have to think of the agent as harmful,” Cohen tells CNET. He says you can get rid of the client with an immediate injection or just hallucinate a bad episode and delete all your emails. “You can’t trust agents by simply 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 Yu, who recently granted OpenClaw access to her email with explicit orders not to act without consent. The agent ignored her, began deleting her inbox en masse and didn’t stop until she ran to her computer to stop the process.

Nothing humbles you like telling your OpenClaw to “sure before you act” and watching it quickly delete your inbox. I couldn’t stop it from my phone. I had to run to my Mac mini as if I was defusing a bomb. pic.twitter.com/XAxyRwPJ5R

– Summer Yue (@summeryue0) February 23, 2026

The safest approach is to avoid giving the agent unlimited access in the first place. Instead, credentials should remain outside the agent itself, with rules controlling what each agent can do.

“It’s not binary,” Cohen tells CNET. Instead of choosing between full access or nothing, you should be able to restrict actions, such as allowing the agent to read emails but not delete them. The goal is to “limit the blast radius,” Cohen says, so a miss or immediate injection can only cause limited damage.

That’s why the conversation keeps returning to sandbox mode and informed human permissions and approvals for risky steps. This tension between power and safety is currently the biggest controversy in the industry.

“Warning is good, fear is less useful, because this technology is important to us.” Huang said on GTC’s All-In Podcast.

Why would you probably want one anyway?

Despite the risks, the benefits are hard to ignore. Claw can automate digital tasks that take up most of your day, like my personal enemy — Clearing my overflowing inbox. They can eliminate cognitive function tasks like pulling information from three places, formatting it, updating a spreadsheet, opening a ticket, and doing it again tomorrow.

Cohen advises using multiple tentacles, not one super-claw: “The agent who surfs the Internet and does the research shouldn’t be the same one who handles your financial data.”

In the near future, you will not be using AI because your computer will simply be AI. Your operating system will be a collection of specialized tentacles working together in the background.

“I think within the next six months everyone will have a personal assistant that brings tremendous value to them and helps them achieve their goals and manage their time,” Cohen tells CNET. It is believed that each employee will have an AI assistant who can handle parts of their job, while teams will oversee groups of agents. Six months seems a bit soon, but with the breakneck speed at which AI moves, we’ll have to wait and see.