Moonshot AI Launches Kimi Work, a Desktop Agent for Localized AI Tasks
Moonshot AI, a prominent Chinese AI startup, has unveiled Kimi Work, a new downloadable desktop agent designed for macOS and Windows. This application aims to bring AI capabilities directly to users’ machines, allowing it to access local files, control web browsers, and execute scheduled tasks. The Beijing-based company announced the product this week, offering free downloads while the app is currently undergoing internal testing.
Bridging the Gap Between Cloud and Local AI
Kimi Work builds upon the foundation laid by Moonshot’s WebBridge extension, which was released a month prior. WebBridge enabled AI agents to control real Chrome or Edge browser sessions locally. Kimi Work expands this concept into a comprehensive desktop product. The core proposition is a departure from the prevailing cloud-based AI model. Typically, users submit prompts to cloud servers, which then utilize sandboxed browsers to perform actions and return results. Kimi Work reverses this paradigm by operating directly on the user’s computer.
Local Access and Enhanced Capabilities
As a locally installed application, Kimi Work gains direct access to a user’s files and can interact with their computer system. This allows it to perform a range of tasks, such as manipulating PDFs, organizing desktop files, retrieving stock data from a browser, compiling reports in HTML format, and sending them via email. The functionality is positioned as a fully developed and integrated solution within the Kimi ecosystem, offering unique features not found in competing tools like OpenClaw or Hermes.
Agent Swarm and Browser Integration
A key innovation in Kimi Work is its Agent Swarm feature. This allows the agent to initiate multiple sub-agents in parallel, potentially up to 300, with each sub-agent handling a distinct part of a larger task. The integration with WebBridge is also crucial, as it grants the agent control of the user’s actual browser through the Chrome DevTools Protocol. This method ensures that logged-in sessions and cookies remain on the user’s machine.
Scheduling, File Access, and Market Data
Kimi Work includes a built-in Cron engine for scheduling tasks based on daily, hourly, or conditional triggers. A “Keep Computer Awake” toggle is available for tasks that need to run overnight. The application also features a local file layer that enables the agent to read mounted folders and run Python scripts in the background. Furthermore, Kimi Work comes with pre-integrated native market data for A-shares, Hong Kong stocks, and U.S. equities, eliminating the need for API setup. Completed research can be directly converted into PowerPoint or Excel formats.
Underlying Technology and Context Window
Internally, Kimi Work is powered by Moonshot’s Kimi K2.6 model, a mixture-of-experts architecture with approximately one trillion parameters. Released on April 20, K2.6 activates around 32 billion parameters per token and boasts a 256K-token context window. This extensive context window is significant, enabling the agent to retain a large amount of information across complex, multi-step workflows without losing track of its initial objectives. The Kimi K2.6 model also served as the base for Cursor’s specialized coding LLM, “Composer 2.”
Cloud Inference and On-Device Options
While Kimi Work’s actions occur locally, the inference for the K2.6 model can still be routed through Moonshot’s cloud API. For users seeking complete on-device inference, the model weights are available on Hugging Face under a Modified MIT License. However, running a trillion-parameter model locally requires substantial hardware resources that are not typically found in standard home user setups.
Privacy Considerations and Safety Measures
The “local” aspect of Kimi Work refers to where actions are executed, not necessarily where the AI model runs. The privacy implications are nuanced, as WebBridge’s control over the user’s logged-in browser means it could potentially access sensitive information like bank accounts, emails, and internal company tools. Researchers have previously highlighted the risk of AI agents exhibiting “blind goal-directedness,” where they execute tasks without fully recognizing potential risks. Moonshot has addressed this by including an “ask before acting” mode, which requires user approval before modifying files or running code. While this is a crucial safety feature, it does not guarantee absolute security.
The Growing Desktop Agent Landscape
The market for desktop AI agents has rapidly expanded. Anthropic’s Claude has offered full desktop computer use since late 2024, while OpenAI released Codex Background Computer Use for macOS in April 2026, enabling parallel desktop sessions. Google’s Gemini computer use, derived from Project Mariner, focuses on browser workflows. Microsoft’s Copilot Studio, introduced in May 2026, targets enterprise automation and integrates models from OpenAI and Anthropic.
User Demand for Flexibility
The trend suggests a user preference for flexibility over being confined to a single provider. Tools like OpenClaw, Hermes, and NanoClaw offer local platforms that can configure AI agents using any LLM via API. Kimi Work distinguishes itself with its local-first design combined with the 300-agent swarm capability. While many competitors operate in cloud sandboxes, limiting their access to real logged-in sessions, or offer desktop control without coordinated parallel agents, Kimi Work aims to provide both. A trade-off is that tasks cease when the laptop is closed, unlike Moonshot’s cloud-based Kimi Claw, which operates 24/7.
Pricing and Availability
Kimi Work is available as a free download, with advanced agent features requiring a paid subscription. Moonshot’s Moderato tier 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 is unlocked at the Allegretto tier ($39/month). The full 300-agent swarm and professional workflows are available in the Allegro ($99/month) and Vivace ($199/month) tiers. Downloads for macOS (Apple Silicon) and Windows are accessible at kimi.com. The internal testing phase means some features may evolve before a broader rollout.
Leave a Reply