Moonshot launches Kimi Work desktop agent

Most AI agents still feel like they live in an isolated lab somewhere in the cloud. You give them a prompt, wait for a remote container to spin up, and watch a virtual browser crawl through tasks. Moonshot AI wants to skip the middleman and live right on your machine.

The Beijing-based startup officially launched Kimi Work, a dedicated desktop agent for macOS and Windows. The app is currently rolling out in an internal testing phase with free downloads, while paid subscriptions start at $19 a month.

Driving your real browser, not a sandbox

Unlike traditional cloud assistants, Kimi Work installs directly on your operating system and executes commands locally. It handles tasks like sifting through messy local folders, analyzing local PDFs, and executing background Python scripts without needing files uploaded to a third-party server.

Its key trick comes from WebBridge, an extension Moonshot shipped a month ago. Using the Chrome DevTools Protocol, Kimi Work takes over your actual, daily Chrome or Edge session. That means your existing logins, active sessions, and cookies remain untouched on your machine, eliminating the need to solve CAPTCHAs or scan QR codes inside a hollow virtual machine.

It also comes equipped with a built-in Cron scheduling engine to trigger workflows hourly, daily, or based on specific conditions. There is even a dedicated toggle to keep your computer awake for overnight batch jobs. For finance workers, native market data for A-shares, Hong Kong stocks, and U.S. equities is pre-integrated out of the box, with outputs converting straight into styled PowerPoint decks or Excel sheets.

A 300-agent swarm backed by K2.6

When a research job gets too big for a single thread, Kimi Work spins up what it calls an Agent Swarm. Locked behind higher subscription tiers, the feature can deploy up to 300 sub-agents in parallel, each carving out a specific slice of the overall workload.

Under the hood, the system is driven by Kimi K2.6, Moonshot's flagship open-weight model released on April 20. It is a roughly one-trillion-parameter mixture-of-experts model that activates about 32 billion parameters per token and boasts a massive 256K context window. Notably, this is the very same architecture Cursor tapped to fine-tune its Composer 2 coding model.

That said, it is critical not to confuse local execution with local inference. While the file reads, clicks, and scripts run on your PC, the heavy LLM inference still routes through Moonshot's cloud API. Although the model weights are hosted on Hugging Face under a Modified MIT License, running a trillion-parameter setup fully on-device is well out of reach for consumer hardware.

Handing an agent keys to your real browser also raises real safety questions. Because it shares your active sessions, the agent can theoretically touch your personal banking, corporate email, and sensitive internal tools. UC Riverside researchers warned in May about "blind goal-directedness," where agents pursue objectives without evaluating peripheral risks. Moonshot includes an "ask before acting" confirmation prompt for file modifications and code execution, but user vigilance remains paramount.

Why it matters

We see Kimi Work as a smart, pragmatic shift in how agentic software reaches everyday workflows.

For knowledge workers, researchers, and financial analysts, the value proposition is immediate. Bypassing cloud sandboxes solves the two biggest headaches in agent automation: authentication friction and local file access. Being able to pull live market data, cross-reference local research PDFs, and generate finished presentations while you sleep is a genuine productivity leap. Starting at $19 a month, it is positioned directly against Western productivity agents, though the premier 300-agent swarm is reserved for pricier plans. Most users should start with the free tier and keep confirmation prompts firmly enabled.

For the broader industry, the battleground has decisively moved from web chat interfaces directly onto user operating systems. Anthropic pioneered computer use late in 2024, OpenAI deployed Codex Background Computer Use on macOS this April, and both Microsoft and Google are pushing deep desktop hooks. Moonshot's decision to leverage CDP on local browsers avoids costly virtualization infrastructure and delivers a snappier experience. However, sitting directly inside a user's logged-in desktop makes security a razor-thin line; a single rogue click or corrupted directory could easily shatter user trust.