Clark

Clark is a computer-use AI agent from Clark Labs, an applied AI lab based in San Francisco. Rather than answering a question and leaving the doing to you, Clark takes a task and carries it out, driving a real computer until there's a finished result to hand back. You describe what you want in a chat, and the agent plans the steps, opens what it needs, and works through them on its own. The idea is to move past the familiar chatbot pattern where the model explains what to do and you still have to go do it yourself. Clark tries to be the part that actually clicks the buttons, so the work happens inside the agent instead of on your screen. That distinction, doing the task rather than describing it, is the line the entire product is built around.

Each task runs inside a dedicated computer that the agent controls, complete with a real browser, a file system, and the ability to schedule work to run later. That environment is what lets Clark handle jobs that take many steps rather than a single reply. It can move through multi-step browser workflows, publish a website, keep watch on a page and check it on a recurring schedule, and gather research into files you can open afterward. Because everything lands as real artifacts, a document, a published page, a saved file, you can inspect exactly what the agent produced instead of taking a summary on faith. When something looks off, the output is right there to review and correct. The scheduling piece matters too, since a job like monitoring a site keeps running on its own without you having to sit and watch it.

The product suits people who want an assistant that finishes things, not one that only describes them. Founders and operators can offload repetitive online chores, researchers can have findings pulled together into files, and anyone comfortable delegating a browser task to software gets back a result they can check. Clark is reachable through the web app at clarkchat.com and on your phone, so a long-running job doesn't chain you to a single device. You hand off a task, step away, and come back to inspectable work. It's aimed at users who are ready to treat an agent as something that owns a workflow from start to finish rather than a place to ask one-off questions, and it rewards giving it jobs that would otherwise eat an afternoon of clicking.

What sets Clark apart is the lab behind it. Clark Labs calls itself the first AI lab run by autonomous AI, meaning the hands-on engineering and research happen in AI loops while the humans contribute taste, intuition, and feedback. That is an unusual operating model, and it has produced a stack of open-source components the agent is built on rather than a thin wrapper over someone else's model. The team has published much of that work openly, which gives you a rare look at the machinery a commercial agent normally keeps hidden. For a young company, that willingness to ship the underlying research in public is a meaningful signal about how seriously it's taking the harder parts of the problem instead of just assembling existing pieces.

Those components each target a hard piece of the agent problem. Clark Browser is a stealth Chromium fork patched so an automated session reads as an ordinary Chrome profile to detector pages, which keeps the agent from getting blocked partway through a task. Clark Hash compresses neural embeddings at a high ratio without any corpus-wide training, a practical answer to storing and searching the large memory an agent accumulates. Clark Air is a model compression pipeline, and Clark Mind is an experimental cognitive architecture that tackles reasoning puzzles through exploration and memory rather than backpropagation. Together they point at the parts of a computer-use agent that usually break, the browser, the memory layer, and the reasoning, and try to fix them at the source instead of papering over them.

Access is subscription-based. The terms describe paid plans that renew automatically until you cancel, so it's worth reading the current tiers on the site before you commit, and canceling ahead of a renewal if you're only trying it out. There isn't a widely published free allowance, so treat this as a paid product with the usual sign-up and billing you'd expect from a hosted agent. What you're paying for is the managed computer, the browser, and the scheduling, all handled for you rather than something you stand up and maintain yourself, which is a real part of the value if you don't want to run that infrastructure.

For anyone watching where computer-use agents are heading, Clark is one of the more opinionated bets in the category. Most tools in this space assemble an agent on top of an existing model and a hosted browser. Clark is building the browser, the memory, and the reasoning underneath its own product and releasing much of it publicly. That makes it interesting even if you're only evaluating the space, and genuinely useful if you have real browser and research work you'd rather hand off and inspect later.