An open-source project with 145,000 GitHub stars is reshaping how the biggest names in AI think about autonomous agents. OpenClaw, the free agent tool that lets AI models actually control your computer - reading files, running commands, browsing the web, sending emails - has gone from viral curiosity to strategic inflection point in under three months.
Now Nvidia, Anthropic, and Perplexity are all scrambling to build on top of it, around it, or against it.
Nvidia Wants to Be the Windows of AI Agents
At GTC 2026, Nvidia announced NemoClaw, an enterprise software stack built on top of the OpenClaw framework. The pitch: single-command installation of Nvidia's Nemotron models with a sandboxed runtime called OpenShell that isolates agent activity behind policy-based guardrails for network access, privacy, and security.
"OpenClaw is the operating system for personal AI," Jensen Huang said during the announcement. "This is the moment the industry has been waiting for."
The CUDA playbook is obvious here. Nvidia spent two decades locking developers into its GPU ecosystem through software tooling. NemoClaw is the same move for agents: give away the framework, own the middleware layer between hardware and business logic, create switching costs through developer loyalty. NemoClaw runs on GeForce RTX PCs, RTX PRO workstations, DGX Station, and DGX Spark machines. No pricing yet, and availability is still vague.
Nvidia also launched the Nemotron Coalition, pulling in Mistral, Perplexity, Cursor, Black Forest Labs, LangChain, and others to co-develop open frontier models. Salesforce, Cisco, Google, and CrowdStrike are in partnership discussions.
Anthropic and Perplexity Take Different Routes
Anthropic's answer is Claude Cowork, an AI agent released in January that works directly with your files and tools. Last week they added Dispatch, which lets you launch Cowork tasks remotely from your phone while the agent runs on your local machine. The angle: keep the agent tethered to your existing workflow rather than letting it roam free across your system.
Perplexity used its first developer conference to position itself as the security-conscious alternative. The company announced an enterprise version of Perplexity Computer, its agent system, plus a preview of Personal Computer - a Mac-native agent with local file access. The implicit message: OpenClaw is powerful but risky; we'll give you the same capabilities with guardrails built in.
The Real Risk Nobody's Pricing In
The security question is the elephant in every demo. OpenClaw agents can send emails, modify files, execute shell commands, and browse the web autonomously. That is an extraordinarily broad attack surface. Reco.ai has already called it "the first major AI agent security crisis of 2026," and the tool's rapid adoption - 20,000 forks and counting - means thousands of developers are running agents with system-level permissions right now.
Nvidia's sandboxed OpenShell and Perplexity's enterprise pitch are both direct responses to this problem. But the tension is fundamental: the more an agent can do, the more useful it is, and the more dangerous it becomes.
For anyone using AI tools daily, the practical takeaway is straightforward. Autonomous agents that control your computer are no longer experimental. They are shipping as products from billion-dollar companies. The gap between "AI assistant that answers questions" and "AI agent that does your job while you sleep" just closed significantly. The companies fighting over this market know the stakes, which is exactly why every major player announced their agent strategy within the same two-week window.