In brief
- Moonshot AI has introduced Kimi Work, a desktop AI assistant for macOS and Windows that scans local files, controls your browser, and runs scheduled tasks
- The software runs on Kimi K2.6, the open-weight model that narrowly beat GPT-5.4 and Claude Opus 4.6.
- Subscriptions begin at $19 per month, with the full 300-agent swarm available only on pricier plans.
Moonshot AI has launched Kimi Work, a downloadable desktop agent for macOS and Windows that runs directly on your device, reads your files, controls your browser, and runs tasks on a predefined schedule. The Beijing-based company—one of China’s recognized AI Tiger startups—unveiled the product this week along with free downloads, with the app currently in closed testing.
The company’s WebBridge extension, rolled out a month ago, already allowed agents to control your actual Chrome or Edge session on your local machine. Kimi Work builds on that concept and evolves it into a complete desktop application.
The idea is straightforward. Most AI tools from AI providers today operate in the cloud. You submit a prompt, a server in a remote data center springs into action, a sandboxed browser performs clicks and scrolls, and you receive a result. Kimi Work takes a different approach.
Kimi Work, since it is a locally installed app on your computer, has direct access to your files and can interact with your system in their place. It can edit your PDFs, tidy up your desktop, pull stock market data from your browser, assemble a report in HTML format, and email it straight to you, and much more.
In practice, it delivers the same kind of functionality that OpenClaw or Hermes attempt to provide, but it is a fully realized and deeply integrated product within the Kimi ecosystem, with unique capabilities that competing options lack.
One of those unique features is Agent Swarm, which allows Kimi Work to spawn multiple sub-agents simultaneously—up to 300 of them, with each one tackling a separate portion of a task.
It also integrates with WebBridge, which gives the agent direct control over your real browser using Chrome DevTools Protocol, the same debugging interface developers rely on; your logged-in sessions and cookies remain entirely on your own machine.
A built-in Cron engine lets you schedule tasks on daily, hourly, or event-based triggers, with a “Keep Computer Awake” option available for jobs that need to run overnight. A local file layer allows the agent to read any folders you choose to mount and execute Python scripts in the background.
The app also ships with native market data for Chinese A-shares, Hong Kong stocks, and U.S. equities built right in—no API configuration is necessary. Completed research reports
It can export directly into PowerPoint or Excel files.
At its core, Kimi Work is powered by Kimi K2.6, according to sources who have had early access. (We haven’t tested it ourselves yet.) K2.6 is a large-scale mixture-of-experts model with approximately one trillion total parameters, released by Moonshot on April 20. With this architecture, only a subset of the model’s parameters are engaged at any given time; K2.6 activates roughly 32 billion parameters per token and supports a 256K-token context window. To clarify: tokens represent the smallest units of text an AI processes, while parameters encompass all the numerical values that encode the model’s knowledge and behavior.
That context window is a big deal: It allows the agent to retain a massive amount of information throughout a lengthy, multi-step task without losing track of its objectives.
If you’re uncertain about Kimi’s capabilities as an AI model, note that it served as the foundation for Cursor—a well-known AI-powered code editor—to train its own specialized large language model for coding, called “Composer 2.”
Setting the record straight first
When Kimi Work says “local,” it describes where tasks are carried out—on your own computer—not necessarily where the AI model itself runs. The K2.6 model can still send its inference requests through Moonshot’s cloud-based API, even though the file operations, browser interactions, and script executions happen right on your device.
For those interested in fully on-device inference, the model weights are publicly available on Hugging Face under a Modified MIT License. However, running a trillion-parameter model locally requires high-end hardware that most consumers don’t typically own.
The privacy situation is more complex than “local means secure.” Since WebBridge operates within your actual browser sessions—where you’re already logged in—it has access to your banking, email, and internal company tools. A UC Riverside study in May highlighted that AI agents frequently execute tasks without understanding the risks involved—a pattern they described as “blind goal-directedness.”
Moonshot has built in a “confirm before acting” mode, which asks for your permission before modifying files or executing code. You should absolutely keep this enabled, but it doesn’t eliminate all risk.
The intensifying competition in agentic AI
The desktop agent market is heating up quickly. Anthropic’s Claude has supported full desktop computer access since late 2024. OpenAI launched Codex Background Computer Use for macOS in April 2026, enabling agents to run in parallel desktop sessions. Google’s Gemini computer use—evolved from Project Mariner—specializes in browser-based workflows. And Microsoft’s Copilot Studio integrated computer-use capabilities in May 2026, targeting enterprise automation with a blend of OpenAI and Anthropic models behind the scenes.
Yet there’s a growing demand among users for flexibility—not wanting to lock themselves into a single AI provider. This is where platforms like OpenClaw, Hermes, and NanoClaw come in. They’re essentially local hubs that let you configure AI agents using any language model accessible through an API.
What sets Kimi Work apart is its local-first approach combined with a 300-agent swarm capability. Most competing solutions either run entirely in a cloud sandbox—meaning they can’t interact with your real logged-in sessions—or offer desktop control without the ability to coordinate parallel agents. Kimi Work delivers both. The catch: if your computer is powered down, the tasks halt. Moonshot’s cloud-based alternative, Kimi Claw, operates around the clock without depending on your machine.
The app itself is free to download, though meaningful agent functionality requires a paid subscription. Moonshot’s Moderato plan starts at $19 per month and includes access to K2.6, Deep Research, and Kimi Code.
The Agent Swarm feature with a limited number of sub-agents unlocks at the Allegretto tier ($39/month). For the full 300-agent swarm and high-volume professional workloads, you’ll need the Allegro ($99/month) or Vivace ($199/month) plans—geared toward those who genuinely need all 300 agents working in parallel.
The app is currently available for download for macOS (Apple Silicon) and Windows at kimi.com. Since it’s still in an internal testing phase, some features may change before the broader public release.
Daily Debrief Newsletter
Start every day with the top news stories right now, plus original features, a podcast, videos and more.
