Kimi WebBridge: AI Agents Take Control of Your Browser While Your Data Stays Private

Moonshot AI, headquartered in Beijing, has launched Kimi WebBridge—a browser extension that enables AI agents to navigate websites just like a human user would. It can search, click, type, scroll, and pull data, all while running locally on your machine. You can download it today from the Chrome Web Store or directly from Kimi’s website.

At the moment, most AI-driven browser automation tools route your information through the cloud, which means your active login sessions and sensitive page content get transmitted to external servers.

Kimi WebBridge takes a different approach. It combines a local background service with the browser extension, and the agent talks to that local service through Chrome DevTools Protocol—the same low-level debugging interface developers rely on. Nothing leaves your computer. Whether it’s your bank account, your email, or your company’s internal dashboards, the agent can work with all of them without Moonshot ever accessing the data.

What Kimi WebBridge actually does

Picture it as giving your AI agent a pair of hands inside your browser. It can navigate to pages, tap buttons, complete forms, capture screenshots to interpret what’s on screen, extract text from web pages, and send results back to whatever AI tool you’re working with. It doesn’t open a separate browser—it works right inside your existing Chrome or Edge window, with all your cookies and active sessions preserved.

For instance, you could tell your agent to search Amazon for mechanical keyboards under $150 with a minimum 4.5-star rating and deliver a ranked comparison list. It would interpret your request and browse Amazon visually rather than relying on API calls. Or you could have it scan LinkedIn job postings across multiple queries and organize them into a spreadsheet, or compare prices for the same item across 10 different retailers and report the best offer.

Any repetitive website task—the kind that eats up 20 minutes of tedious manual effort—gets reduced to a single sentence.

Kimi WebBridge officially supports Kimi Code CLI, Claude Code, Cursor, Codex, and Hermes. So it isn’t tied to the Kimi ecosystem. If you’re already using a different AI coding or agent tool, WebBridge integrates with it seamlessly.

Kimi: The Chinese AI model US tech giants quietly rely on

WebBridge is driven by Moonshot AI’s Kimi model family—developed in China—and the underlying technology is more powerful than most Western users appreciate. Kimi K2 debuted in July 2025 as a 1-trillion-parameter, open-source mixture-of-experts model, claiming the top spot among open-source models and fifth place overall on the LMSYS Arena leaderboard. (Parameters define a model’s scope of knowledge—higher is typically better, though not universally so.)

The newest release, K2.6, launched in April 2026, now achieves a 58.6% score on SWE-Bench Pro—a benchmark that evaluates real-world software engineering skills on actual GitHub issues—edging past GPT-5.4 at 57.7% and Claude Opus 4.6 at 53.4%.

If Kimi hasn’t crossed your radar before, here’s why it’s making headlines in 2026: the Cursor controversy. On March 19, the $50 billion coding AI company Cursor unveiled its Composer 2 model, promoting it as “frontier-level proprietary coding intelligence” developed through “continued pretraining” and reinforcement learning.

The announcement survived fewer than 24 hours before a developer named Fynn captured API traffic and uncovered a model identifier: kimi-k2p5-rl-0317-s515-fast. Elon Musk responded with three words: “Yeah, it’s Kimi 2.5.” Moonshot’s head of pretraining conducted a tokenizer analysis. Perfect match, confirmed.

Lee Robinson, VP of Developer Education at Cursor, quickly recognized the open-source foundation and clarified that about 75% of the compute was dedicated to Cursor’s own training pipeline. Aman Sanger, co-founder of Cursor, described the omission as “a miss from the start.” Moonshot responded graciously—publicly congratulating Cursor on its launch.

Who else is in this space

Browser automation for AI agents is becoming increasingly competitive. Anthropic’s computer use feature enables Claude to interact directly with desktops. OpenAI’s Operator and ChatGPT Atlas offer similar capabilities through a cloud-hosted service. Google is experimenting with DeepMind-powered agents, and Perplexity has its own Comet Browser.

What sets Kimi WebBridge apart—at least for now—is its local-first design. Cloud-based browser agents are convenient, but they require sending your browsing activity through a third-party server. When it comes to personal accounts or sensitive information, that’s a significant trade-off to consider.

To get started, the quickest route to WebBridge is through the official setup page at kimi.com/features/webbridge, which guides you through each step in sequence.

Step 1: Download the Kimi Desktop App. The extension requires Kimi Claw Desktop, which runs on your local machine. Mac users can download it directly from the setup page; Windows users can install it via PowerShell by running:

irm | iex

Step 2: Install the browser extension from the Chrome Web Store, or manually via the setup page.

Step 3: Open the Kimi Desktop App, locate Kimi Claw in the left sidebar, create a new Claw, and choose “On my computer” to set it up as a local agent. Then give it a task—for example, “Browse Amazon for a mechanical keyboard under $150 with 4.5+ stars”—and let it get to work.

For other AI agents like Claude Code, Cursor, or Codex, the setup page provides a connection command that you paste into your agent, automatically linking it to the WebBridge service.

If the extension appears disconnected, simply resend the connection command in Kimi Claw Desktop and restart the app. The most frequent issue is that the local service wasn’t running before the extension attempted to connect.

Moonshot states that K2.6 can support up to 300 parallel sub-agents executing across 4,000 coordinated steps at the same time—the underlying architecture that WebBridge leverages when tackling complex, multi-step browser tasks.