AI agents · OpenClaw · self-hosting · automation

Quick Answer

Claude vs ChatGPT vs Kitesurf: Agent Browsers 2026

Published:

The Short Answer

Claude Cowork browserChatGPT Work browserCloudflare Kitesurf
FormSide panel in desktop appHeadless Chrome in Work environmentAgent-first browser runtime
Where it runsLocally, via desktop appServer-side (Work Cloud) or local (Work Local)Cloudflare edge
Needs your machine on✅ Yes❌ No (Cloud tier)❌ No
Credential importChrome/Edge/Firefox; banking + email excluded by defaultWithin Work environmentProgrammatic
Best forAttended research, multi-site gatheringUnattended long tasks + code executionProgrammatic agent traffic at scale
Injection exposureHighHighest (code exec + filesystem + deploy)Contained by design

Last verified: August 31, 2026.

The Architectural Split That Actually Matters

Ignore the feature lists for a moment. There is exactly one question that separates these three products: where does the browser process live?

Claude Cowork: on your machine. Anthropic ships the browser inside the desktop app, launched in late August 2026. Pages are fetched from your network and your IP. Geo-restricted content and IP-allowlisted internal tools remain reachable. The cost is that the desktop app must be running — even for cloud Cowork sessions — so nothing runs while your laptop is shut.

ChatGPT Work: on OpenAI’s servers (Cloud tier) or your desktop (Local tier). Work Cloud runs through chatgpt.com with a genuinely headless Chrome; Work Local is a desktop app. The Cloud tier is the only one of the three that can browse for hours while you sleep.

Kitesurf: on Cloudflare’s edge. It is not a browser a human uses at all — it is a browser runtime designed for agent traffic, addressed programmatically.

Every other difference downstream of this is a consequence of that choice.

Capability, Honestly Ranked

ChatGPT Work is the most capable — and the most alarming

Simon Willison’s hands-on breakdown, published in late August 2026, catalogued what ships in the Work environment: internet-connected code execution, a headless Chrome, a persistent /workspace/scratch filesystem, ChatGPT Sites deployment onto Cloudflare Workers, sub-agents, and scheduled automations.

His verdict was that the combination hits the full lethal trifecta — private data, exposure to untrusted content, and an exfiltration path — and he criticised OpenAI’s documentation for emphasising use cases over technical specification.

He is right, and the deployment primitive is the part people underrate. An agent that can read a hostile web page and publish a Cloudflare Worker has a first-class exfiltration channel that does not look like exfiltration. It looks like the product working.

Claude Cowork is the most usable

Zero setup, included in Pro/Max/Team/Enterprise, visible in a side panel you can watch. Anthropic’s positioning — “a lot of web tasks don’t need your browser, just a browser” — is aimed at cutting the Chrome-extension dependency, and by extension Google.

The visibility is a real safety property, not a consolation prize. An attended agent whose actions render in front of you fails loudly. A headless agent fails silently at 3am.

Kitesurf is the only one designed for the problem

Kitesurf starts from the premise that agents are not humans and should not pretend to be. That reframing is what lets it contain injection risk structurally rather than by policy — there is no imported credential vault, no user filesystem, no daily-driver profile to compromise.

The trade is obvious: it is infrastructure. You write code. It solves nothing for a knowledge worker who wants a research task done this afternoon.

The Credential Question

This is where the products genuinely diverge on judgement, not just architecture.

Anthropic lets you import saved logins from Chrome, Edge and Firefox but excludes banking and email by default. Keep that exclusion. Email is the master key to almost every account through password reset — an agent with mailbox access and an injection vulnerability is an account-takeover engine, not a productivity tool.

The general rule across all three: assume anything the agent can see, it can exfiltrate. Scope credentials to that assumption. Use a dedicated browser profile for agent work. Never point an agent at your identity provider.

Which To Pick

Pick Claude Cowork’s browser if you have a Claude subscription, want research and multi-site data gathering done while you watch, and value setup cost near zero.

Pick ChatGPT Work if the task genuinely needs code execution plus browsing plus persistence over hours — and you are willing to run it against non-sensitive accounts in a blast-radius-limited environment.

Pick Kitesurf if you are building a product where agents browse at volume and you want the runtime to be honest about that.

Pick none of them for anything touching payments, email, or production credentials. That boundary has not moved in 2026.

Sources