The Open-Source AI Agent Gold Rush: How OpenClaw Went From Zero to a Million Searches in 60 Days

Something broke loose in the Rising Trends database at the start of 2026. The term "OpenClaw" had sat at roughly 10 monthly searches for five straight years. In January it hit 60,500. In February it hit 1,000,000. It was not alone. A whole cluster of related queries (Hermes agent, Claude Cowork, ClawHub, cmux, Copilot Cowork) broke out inside the same eight-week window.

This is the moment "AI agent" stopped being a research-lab phrase and became something millions of people typed into a search bar. Below we walk through what actually happened, the taxonomy of these new agents, why the open-source ones spread fastest, the security incident that already followed, and the economy forming around all of it.

Let's get into it.

What we're seeing in the Rising Trends data

Here is the monthly search volume for "OpenClaw" over the past year, pulled directly from our database. It is one of the cleanest hockey-stick curves we have ever recorded.

The plain "OpenClaw" curve is the headline, but the real story is the cluster. When we pulled every agent-related breakout query from our database, the landscape looks like this.

A few things jump out. OpenClaw dominates the cluster at 450,000 monthly searches, but the more interesting signal is the breadth. Eight distinct products, from scrappy open-source projects to features shipped by Microsoft and Anthropic, all broke out together. When that many adjacent terms spike inside one quarter, you are not looking at a single viral repo. You are looking at a category being born.

What an "AI agent" actually means here

For two years, "AI" meant a chatbot: you type, it types back. The 2026 breakout is about something different. An agent does not just respond, it acts. It reads and writes files on your machine, runs commands, browses the web, sends messages, and chains those steps together to complete a goal without a human approving each move.

OpenClaw, per our database, is "an open-source, autonomous AI agent framework that runs locally and connects to messaging platforms for tasks like email management, web browsing, and workflow automation." The key words are autonomous, local, and open-source. That combination is exactly why it spread the way it did. It runs on your own hardware, anyone can read and extend the code, and it does real work instead of just talking about it.

The shift is the same one we documented across the broader space in our AI agents trends 2026 report: the frontier moved from "models that answer questions" to "systems that take actions." Search behavior is now catching up to that shift in real time.

The taxonomy: four kinds of agent in one cluster

The cluster looks chaotic until you sort it by what each tool is actually for. There are four distinct buckets.

Local autonomous agents. This is the OpenClaw lane, and it is the one driving the headline numbers. OpenClaw was built in late 2025 by Peter Steinberger, the Austrian developer who previously founded PSPDFKit, and it runs as a 24/7 personal assistant on your own machine that plugs into your messaging apps. Hermes, from Nous Research, is the other heavyweight here at 201,000 searches, an MIT-licensed assistant launched in February 2026 with persistent memory and a built-in learning loop that lets it write its own reusable skills. Both are community-driven, both run locally, and both prioritize privacy and extensibility over a polished hosted product.

Enterprise coworkers. This is the lane the large labs are funding. Claude Cowork (246,000 searches) is Anthropic's feature that turns Claude into a hands-on collaborator with local file access and workflow automation; it reached general availability in April 2026 for paying subscribers on macOS and Windows. Copilot Cowork is Microsoft's analogous product, built with Anthropic, for autonomously executing multi-step enterprise tasks like spreadsheets and reports. The framing here is explicitly the office: an agent that sits beside knowledge workers rather than on a hobbyist's laptop.

Agent-orchestration tooling. Once people run agents all day, they need to manage them. cmux (27,100 searches) is a native macOS terminal app for running multiple coding agents (Claude Code, Codex, Gemini) in parallel with better visibility into what each one is doing. This bucket barely existed six months ago. Its emergence is a tell: you only build dashboards for a thing once that thing is everywhere.

Multi-model systems. Perplexity Computer (60,500 searches) stitches several models together into long autonomous workflows for research and data processing. It points at where this is heading, agents that route work across many models rather than depending on one.

The four buckets share one trait: every term is flagged EXPONENTIAL in our pace classification. That is rare. Usually a category has a leader and a long tail of flat also-rans. Here the entire field is accelerating at once.

Why the open-source ones won the search war

The most instructive part of this data is that the fastest-growing agents were not the best-funded ones. OpenClaw and Hermes, both open-source and community-run, out-grew several commercial products with real marketing budgets behind them.

Three forces explain it. First, distribution. Open-source agents spread through GitHub stars, forks, and developer word of mouth, which compounds far faster than a marketing funnel. OpenClaw crossed 302,000 GitHub stars by April 2026 to become the fastest-growing open-source project in the platform's history, and Hermes pulled in more than 180,000 stars in under four months. "openclaw github" was itself a 27,100-search query, people searching specifically to find and clone the repo.

Second, trust through transparency. An agent that can read your files and send messages on your behalf is asking for an extraordinary level of access. Being able to inspect the code is not a nice-to-have for that kind of software; it is the entire basis for trusting it. Local execution reinforces this, your data never has to leave your machine.

Third, extensibility. Hermes can author its own skills. OpenClaw has a marketplace of community-built ones. That turns each tool into a platform, and platforms grow faster than products because the users build the features.

The security bill came due immediately

Here is where the story stops being a pure growth narrative. The same openness that fueled adoption created an attack surface, and it was exploited almost at once.

ClawHub, the official skill marketplace for OpenClaw (27,100 searches in its own right), became the agent era's first mass supply-chain incident. Security firm Koi Security audited the registry and, in its ClawHavoc research, found 341 malicious skills, 335 of them from a single coordinated campaign. The poisoned skills masqueraded as legitimate tools and crypto utilities while delivering information-stealing malware (the Atomic macOS Stealer and Windows infostealers) that harvested LLM API keys, SSH keys, browser-saved passwords, and crypto wallet data. The mechanics are worth understanding because they will repeat. A "skill" is third-party code that an autonomous agent installs and runs with the same permissions as the agent itself, which for a local agent means broad access to files, credentials, and the network. A poisoned skill is not a chatbot giving a bad answer. It is hostile code running on your machine with your agent's full authority. The only barrier to publishing one of these skills, per Koi's analysis, was a GitHub account that was at least a week old.

This is the unavoidable tension of the whole category. The features that make these agents useful (autonomy, local access, community extensions) are exactly the features that make a compromise catastrophic. Anyone evaluating an agent right now should be treating skill marketplaces with the same caution they would apply to installing random browser extensions or npm packages from unknown authors, because structurally that is what they are.

The economy forming around agents

A category growing this fast pulls an economy along with it, and you can already see the shape of it in the cluster.

There is the orchestration layer (cmux and tools like it) selling visibility and control over fleets of agents. There is the enterprise layer, where Anthropic's Claude Cowork and Microsoft's Copilot Cowork are racing to make the agent a standard fixture of office software, the way the spreadsheet became one. There is the skill and marketplace layer, monetizing community-built capabilities (and, as ClawHub shows, now needing to fund a security and review function on top). And there is a fast-growing services layer of consultants and shops that set up and babysit agents for businesses that want the productivity without the operational risk.

The commercial center of gravity is the enterprise coworker. That is where the budgets are, and the "Cowork" naming convention adopted by both Anthropic and Microsoft is a deliberate attempt to define the agent as a colleague rather than a tool. Whoever wins the vocabulary tends to win the category.

Where this is heading

Three things look likely through the rest of 2026.

The cluster will keep spawning variants. "Cowork" is already becoming a generic suffix, the way "-maxxing" did in a completely different corner of culture (a pattern we broke down in our maxxing trend explainer). Expect more "[brand] agent" and "[brand] cowork" terms to surface in next quarter's search data.

Security becomes a first-class product category. The ClawHub incident will not be the last. Skill vetting, agent sandboxing, and permission management are going to turn into their own products and their own search trends. This connects directly to the themes in our cybersecurity trends 2026 coverage, the attack surface is moving from the network to the agent.

The open-source and enterprise lanes converge. The local hobbyist agents will get enterprise-grade controls, and the enterprise coworkers will get more open and extensible to compete with community ecosystems. The interesting companies will be the ones that sit in the middle.

The signal worth watching is the orchestration layer. Tools like cmux exist only because people now run many agents at once. When the management tooling for a technology starts trending on its own, the underlying behavior has already gone mainstream. By that measure, the agent era is not coming. According to our data, it arrived in February.

---

Want to spot shifts like this before they hit the headlines? Read our guide on how to identify market trends, see the live OpenClaw trend page, or browse what is breaking out right now on the Rising Trends dashboard.