Summary
- Moonshot AI has introduced Kimi Work, a desktop AI assistant for macOS and Windows that can access local files, operate your browser, and perform scheduled tasks.
- The application is powered by the Kimi K2.6 model, which surpasses GPT-5.4 and Claude Opus 4.6.
- Subscription plans begin at $19 per month, but access to the complete suite of 300 agents is available only at higher tiers.
Moonshot AI has unveiled Kimi Work, a downloadable AI agent for macOS and Windows that operates directly on your computer, capable of reading files, controlling your web browser, and executing tasks based on a schedule. The Beijing-based firm—recognized as one of China's emerging AI startups—announced this product earlier this week, offering free downloads while the app is currently undergoing internal testing.
Previously, the company released the WebBridge extension, which allowed agents to control your real browser sessions on Chrome or Edge. Kimi Work expands this concept into a comprehensive desktop application.
The concept is straightforward: Unlike most AI tools that operate in the cloud—where you send a prompt and a server processes it—Kimi Work functions locally.
As a locally installed application, Kimi Work can access your files and interact with your computer. It can perform tasks like editing PDFs, organizing your desktop, extracting stock information from your browser, generating reports in HTML, and emailing them to you.
It offers capabilities similar to those of OpenClaw or Hermes but is fully integrated within the Kimi ecosystem, featuring unique functionalities not found in other options.
One notable feature is the Agent Swarm, which allows Kimi Work to deploy multiple sub-agents simultaneously—up to 300—each handling a different aspect of a task.
Additionally, it integrates with WebBridge, granting the agent control over your actual browser using the Chrome DevTools Protocol, which developers use for debugging; your logged-in sessions and cookies remain secure on your device.
A built-in scheduling engine allows tasks to be set based on daily, hourly, or conditional triggers, with a "Keep Computer Awake" feature for overnight tasks. It also includes a local file layer that enables the agent to access mounted folders and execute Python scripts in the background.
The application comes pre-equipped with market data for A-shares, Hong Kong stocks, and U.S. equities, eliminating the need for API setup. Completed research can be directly converted to PowerPoint or Excel formats.
Kimi Work operates on the Kimi K2.6 model, which is reported to have around one trillion parameters and was launched on April 20. This mixture-of-experts model activates only a portion of its parameters at any given time, engaging approximately 32 billion per token while utilizing a 256K-token context window. (For context, tokens represent the smallest units of information an AI can process, while parameters are the numerical values that encapsulate a model's knowledge and characteristics.)
The significance of the context window is that it enables the agent to retain substantial information throughout lengthy and multi-step workflows without losing track of its initial tasks.
If there's skepticism about Kimi's AI capabilities, it's worth noting that this model served as the foundation for the widely used AI code editor Cursor to refine its own specialized coding language model, "Composer 2."
Addressing a Common Misunderstanding
The term "local" in Kimi Work pertains to where actions are executed—on your machine—rather than the location of the AI model's operation. The K2.6 model can still process through Moonshot's cloud API, even while file access, browser interactions, and Python executions occur locally.
For those seeking complete on-device inference, the model weights can be found on Hugging Face under a Modified MIT License, though a trillion-parameter model requires robust hardware that most home users may not possess.
The privacy considerations are also more complex than simply equating local with safety. Since WebBridge operates your actual logged-in browser, it can potentially access sensitive information such as bank accounts, emails, and internal company tools. Researchers at UC Riverside cautioned in May that AI agents often perform tasks without recognizing the potential risks involved, a phenomenon they termed "blind goal-directedness."
Moonshot has included an "ask before acting" feature that requires user approval before any modifications to files or code execution. While this is a prudent default setting, it does not guarantee complete safety.
The Intensifying Competition in Agentic AI
The market for desktop agents has rapidly become competitive. Anthropic's Claude has been offering full desktop functionality since late 2024. OpenAI introduced Codex Background Computer Use for macOS in April 2026, allowing agents to run in parallel desktop sessions. Google's Gemini, evolving from Project Mariner, focuses on browser-based workflows. Microsoft's Copilot Studio added desktop capabilities in May 2026, catering to enterprise automation and leveraging both OpenAI and Anthropic models.
However, users appear to prefer flexibility rather than being tied to a single provider. This has led to the emergence of tools like OpenClaw, Hermes, and NanoClaw, which serve as local platforms enabling users to configure AI agents with any LLM via API.
What distinguishes Kimi Work is its local-first approach combined with the ability to utilize a 300-agent swarm. Many competitors either operate solely within a cloud sandbox—thus unable to access real logged-in sessions—or provide desktop control without coordinated parallel agents. Kimi Work achieves both functionalities. The tradeoff is that when your laptop is shut, tasks cease. In contrast, Moonshot's cloud product, Kimi Claw, can operate continuously without requiring your machine to be on.
The application is available for free download, but significant agent features necessitate a subscription. Moonshot's Moderato tier starts at $19 per month, offering access to K2.6, Deep Research, and Kimi Code.
The limited sub-agent feature is unlocked at the Allegretto tier ($39/month), while the full 300-agent swarm and advanced professional workflows require the Allegro ($99/month) or Vivace ($199/month) tiers for users who truly need all 300 agents.
Downloads for macOS (Apple Silicon) and Windows are currently available at kimi.com, with the internal testing phase indicating that some features may still be subject to change before a broader release.