Moonshot AI’s Kimi Work Brings 300 AI Agents to Your Desktop

Moonshot AI has released Kimi Work, a downloadable desktop agent for macOS and Windows that lives on your machine, reads your files, drives your browser, and runs jobs on a schedule. The Beijing-based company—one of China’s so-called AI Tiger startups—announced the product this week alongside free downloads, with the app currently in internal testing.

The company’s WebBridge extension, released a month ago, already let agents drive your real Chrome or Edge session locally. Kimi Work takes that idea and turns it into a full desktop product.

The pitch is simple. Most AI tools from AI providers today live in the cloud. You send a prompt, a server somewhere spins up, a sandboxed browser clicks around, and you get a result. Kimi Work does the opposite.

Kimi Work, being a local app installed on your computer, has access to your local files and is able to interact with your computer instead. It can manipulate your PDFs, organize your desktop, pull stock data from your browser, compile a report in HTML and email it to you, etc.

In practice it’s the equivalent of what OpenClaw or Hermes try to do, but fully developed and integrated in the Kimi ecosystem with special features that other alternatives don’t offer.

One of those things is Agent Swarm, which lets Kimi Work spin up multiple sub-agents in parallel—up to 300 of them, each handling a different slice of a task.

It’s also integrated with WebBridge, which hands the agent control of your real browser using Chrome DevTools Protocol, the same interface developers use for debugging; your logged-in sessions and cookies stay on your machine.

A built-in Cron engine schedules tasks on daily, hourly, or conditional triggers, with a “Keep Computer Awake” toggle for overnight jobs. And a local file layer lets the agent read folders you mount and run Python in the background.

The app also ships with native market data for A-shares, Hong Kong stocks, and U.S. equities pre-integrated—no API setup required. Finished research converts directly to PowerPoint or Excel.

Under the hood, Kimi Work runs on Kimi K2.6, according to those who have been able to take a peek. (We’ve yet to test this ourselves). K2.6 is a roughly one-trillion-parameter mixture-of-experts model Moonshot released on April 20. Mixture of experts is an architecture that keeps only a fraction of its parameters active at any moment; K2.6 activates around 32 billion per token while carrying a 256K-token context window. (For reference, tokens are the smallest amount of information an AI can process while parameters are all the numerical values that store all the knowledge and traits a model has.)

That context window matters: It means the agent can hold an enormous amount of information in mind across a long, multi-step workflow without forgetting what it started on.

If you have doubts about Kimi’s quality as an AI model, consider that this was the base used by the popular AI code editor Cursor to fine tune its own specialized coding large language model, “Composer 2.”

The misconception worth clearing up first

“Local” in Kimi Work refers to where the actions happen—your machine—not necessarily where the AI model runs. The K2.6 model inference can still route through Moonshot’s API in the cloud, even while the file reads, browser clicks, and Python executions happen locally.

If you want full on-device inference, the weights are available on Hugging Face under a Modified MIT License—but a trillion-parameter model demands serious hardware most home users don’t have.

The privacy picture is also more nuanced than “local equals safe.” Because WebBridge drives your actual logged-in browser, it can touch your bank account, your email, and your company’s internal tools. Researchers at UC Riverside warned in May that AI agents often carry out tasks without recognizing when their actions are risky—a behavior they called “blind goal-directedness.”

Moonshot includes an “ask before acting” mode that requires your approval before any file is modified or code is run. That’s the right default to leave on, but it’s still far from being 100% safe.

The brewing war in the agentic AI era

The desktop agent race has gotten crowded fast. Anthropic’s Claude has offered full-desktop computer use since late 2024. OpenAI shipped Codex Background Computer Use for macOS in April 2026, running agents in parallel desktop sessions. Google’s Gemini computer use—descended from Project Mariner—focuses on browser workflows. Microsoft’s Copilot Studio added computer-use in May 2026, aimed at enterprise automation and using both OpenAI and Anthropic models under the hood.

But it seems like instead of being constrained to a single provider, users want some flexibility. This is where tools like OpenClaw, Hermes, or NanoClaw enter in the game. They are basically local platforms that configure your AI agents using any LLM over API.

The difference with Kimi Work is the local-first design paired with a 300-agent swarm. Most competitors either run everything in a cloud sandbox—which can’t touch real logged-in sessions—or offer desktop control without coordinated parallel agents. Kimi Work does both. The tradeoff is this: When your laptop is closed, tasks stop. Moonshot’s own cloud product, Kimi Claw, runs 24/7 without your machine.

The app is a free download. Meaningful agent features require a paid plan. Moonshot’s Moderato tier starts at $19/month and includes K2.6, Deep Research, and Kimi Code access.

The Agent Swarm with limited sub-agents unlocks at Allegretto ($39/month), while the full 300-agent swarm and highest-volume professional workflows require the Allegro ($99/month) or Vivace ($199/month) tiers for those who really need 300 agents.

Downloads for macOS (Apple Silicon) and Windows are live at kimi.com, with the internal testing phase meaning some features may still shift before a wider rollout.