Kimi’s work for Moonshot AI brings 300 artificial intelligence agents to your desktop

Moonshot AI released chemical action, A downloadable desktop agent for macOS and Windows that sits on your device, reads your files, launches your browser, and manages tasks on a schedule. The Beijing-based company, one of China’s so-called AI Tiger startups, announced the product this week alongside free downloads, and the app is currently undergoing internal testing.

The company’s WebBridge extension, Released A month ago, I already allowed agents to drive a real Chrome or Edge session locally. Kimi Work takes this 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 submit a prompt, a server runs somewhere, the Sandboxed Browser clicks around, and you get a result. ChemiWork does the opposite.

Kimi Work, as a local application installed on your computer, has access to your local files and can interact with your computer instead. It can handle your PDF files, organize your desktop, pull inventory data from your browser, compile an HTML report and email it to you, etc.

In practice, it is equivalent to what OpenClaw or Hermes are trying to do, but it has been fully developed and integrated into the Kimi ecosystem with special features that other alternatives do not provide.

One such thing is Agent Swarm, which allows Kimi Work to spin multiple sub-agents in parallel – up to 300 of them, each handling a different slice of the mission.

It’s also integrated with WebBridge, which handles proxy control in your real browser using the Chrome DevTools protocol, the same interface that software developers use to debug; The sessions and cookies you’re signed in with remain on your device.

The built-in Cron engine schedules tasks to daily, hourly, or conditional triggers, with a “Keep Computer Awake” toggle for nightly tasks. The local file layer allows the agent to read the folders you have mounted and run Python in the background.

The app also comes with local market data for A-shares, Hong Kong stocks and US stocks pre-integrated – no API setup required. The final research is converted directly into PowerPoint or Excel.

Under the hood, Kimmy Work is running around Such as K2.6depending on who they were Able to peek. (We haven’t tested this ourselves yet.) K2.6 is a Moonshot model of approximately one trillion parameters and was released on April 20. An expert mix is ​​a structure that keeps only a portion of its parameters active at any given moment; K2.6 activates about 32 billion per token while holding a context window of 256 thousand tokens. (For reference, tokens are the smallest amount of information an AI can process while parameters are all numeric values ​​that store all the knowledge and attributes the model has.)

This window of context is important: it means that the agent can hold a huge amount of information in mind across a long, multi-step workflow without forgetting what they started with.

If you have doubts about Kimi’s quality as an AI model, keep in mind that this is the rule that popular AI code editor Cursor used to fine-tune its own large, specialized programming language model, “Composer 2.”

The misconception is worth clarifying first

“Local” in Kimi Work refers to where the actions happen – your device – and not necessarily where the AI ​​model is running. K2.6 model inference can still route via the Moonshot API in the cloud, even during file reads, browser clicks, and Python operations performed locally.

If you want full inference on the device, the weights are available on Hugging Face under a modified MIT license – but the trillion-parameter model requires serious hardware that most home users don’t have.

The picture of privacy is also more nuanced than the phrase “locality equals security.” Since WebBridge runs the actual browser you’re signed in to, it can touch your bank account, email, and your company’s internal tools. Researchers in UC Riverside warned In May, AI agents often undertake tasks without realizing when their actions are risky, a behavior they call “blind target orientation.”

Moonshot includes a “Ask before you act” mode that requires your approval before modifying any file or running code. This is the correct default to leave, but it’s still far from 100% safe.

Brewing war in the age of agentic artificial intelligence

The desktop proxies race is quickly becoming crowded. Anthropic’s Claude has offered full desktop use since late 2024. OpenAI shipped Codex Backup Computer Use for macOS in April 2026, running agents in parallel desktop sessions. Using Google’s Gemini computer, which is descended from Project Mariner, focuses on browser workflow. Microsoft’s Copilot Studio added PC usage in May 2026, aiming to automate organizations and use OpenAI and Anthropic models under the hood.

But it seems that instead of being limited to one provider, users want some flexibility. This is where tools like OpenClaw, Hermes, or NanoClaw come into the game. They are basically native platforms that configure your AI agents using any LLM via API.

The difference with Kimi Work is the local-first design combined with a squadron of 300 agents. Most competitors either run everything in a cloud sandbox — which can’t touch real login sessions — or offer desktop control without coordinated parallel agents. ChemiWork does both. The trade-off is this: When your laptop shuts down, tasks stop. Moonshot’s cloud product, Kimi Claw, runs 24/7 without your device.

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

Agent Swarm unlocks with a limited number of sub-agents in Allegretto ($39 per month), while a full agent swarm of 300 agents and a higher-volume Pro workflow requires Allegro ($99 per month) or Vivace ($199 per month) tiers for those truly Need 300 agents.

Downloads for macOS (Apple Silicon) and Windows are available directly at chemi.com, With an internal testing phase which means some features may change before a wider rollout.