Moonshot’s Kimi WebBridge runs AI agents locally in Chrome

Moonshot AI released Kimi WebBridge, a Chrome extension and a local background service that lets AI agents operate inside Chrome and Edge through the Chrome DevTools Protocol. The company says browsing activity, form entries and screenshots remain on the user’s device and do not pass through Moonshot servers. The extension is available on the Chrome Web Store and from Kimi’s official setup page.

The extension pairs with the Kimi Desktop App, which runs a local service called Kimi Claw. Agents connected to the local service can open pages, click, type, scroll, capture screenshots and extract text from the browser window the user already has open. Because the agent communicates with a local process, cookies and logged-in sessions remain on the user’s machine. Moonshot’s documentation states the company does not receive page content or credentials.

Installation requires the Kimi Desktop App and the browser extension. Mac users can download the app from the setup page; Windows users can install via a PowerShell command provided by Moonshot. After launching the desktop app and starting Kimi Claw, users add a new Claw and select “On my computer” to deploy a local agent. For third-party agents such as Claude Code, Cursor, Codex and Hermes, the setup page provides a connection command that links the agent to the local WebBridge service. Moonshot notes a common support issue is the extension attempting to connect before the local service is running; restarting the app or resending the connection command normally resolves the problem.

Moonshot describes WebBridge as agent-agnostic and lists support for Kimi Code CLI, Claude Code, Cursor, Codex and Hermes. The company presents typical use cases as repetitive site workflows such as price comparisons, scraping job listings into a spreadsheet or visually searching retail pages for products that meet specific criteria.

WebBridge runs on Moonshot’s Kimi model family. Moonshot reported that its latest model, K2.6, released in April 2026, scored 58.6% on SWE-Bench Pro, a benchmark that measures software engineering tasks on real GitHub issues. The company’s data places that score ahead of GPT-5.4 at 57.7% and Claude Opus 4.6 at 53.4%. Kimi K2 debuted in July 2025 as a 1-trillion-parameter, open-source mixture-of-experts model. Moonshot says K2.6 can coordinate up to 300 parallel sub-agents executing across 4,000 steps for complex, multi-step browser tasks.

In March 2026, a third-party developer found a model identifier tied to Kimi in API traffic after the company released a product built on that model family without initially disclosing the open-source base. Lee Robinson, the other company’s vice president of developer education, acknowledged the open-source base and added that roughly 75% of the compute went into the company’s own training pipeline. Aman Sanger, a co-founder, described the omission as ‘a miss from the start.’ Moonshot posted a public message congratulating the company on its launch.

Companies including Anthropic, OpenAI, Google and Perplexity offer tools that let models interact with web content or desktops, and many of those services route browsing activity through cloud servers. Moonshot positions WebBridge as a local-first option intended to keep page interactions and credentials on users’ machines.

The Kimi setup page includes step-by-step installation instructions and troubleshooting pointers. Moonshot states that running the agent locally reduces exposure of sensitive data such as bank or work account information.