Moonshot AI's Kimi Work Brings 300 AI Brokers to Your Desktop – Decrypt

Moonshot AI has launched Kimi Work, a downloadable desktop agent for macOS and Home windows that lives in your machine, reads your recordsdata, drives your browser, and runs jobs on a schedule. The Beijing-based firm—certainly one of China’s so-called AI Tiger startups—introduced the product this week alongside free downloads, with the app at the moment in inside testing.

The corporate’s WebBridge extension, launched a month in the past, already let brokers drive your actual Chrome or Edge session regionally. Kimi Work takes that concept and turns it right into a full desktop product.

The pitch is straightforward. Most AI instruments from AI suppliers right this moment stay within the cloud. You ship a immediate, a server someplace spins up, a sandboxed browser clicks round, and also you get a end result. Kimi Work does the alternative.

Kimi Work, being an area app put in in your pc, has entry to your native recordsdata and is ready to work together along with your pc as an alternative. It may possibly manipulate your PDFs, manage your desktop, pull inventory knowledge out of your browser, compile a report in HTML and e-mail it to you, and many others.

In follow it’s the equal of what OpenClaw or Hermes attempt to do, however absolutely developed and built-in within the Kimi ecosystem with particular options that different options don’t provide.

A kind of issues is Agent Swarm, which lets Kimi Work spin up a number of sub-agents in parallel—as much as 300 of them, every dealing with a distinct slice of a job.

It’s additionally built-in with WebBridge, which palms the agent management of your actual browser utilizing Chrome DevTools Protocol, the identical interface builders use for debugging; your logged-in periods and cookies keep in your machine.

A built-in Cron engine schedules duties on every day, hourly, or conditional triggers, with a “Preserve Laptop Awake” toggle for in a single day jobs. And an area file layer lets the agent learn folders you mount and run Python within the background.

The app additionally ships with native market knowledge for A-shares, Hong Kong shares, and U.S. equities pre-integrated—no API setup required. Completed analysis converts on to PowerPoint or Excel.

Underneath the hood, Kimi Work runs on Kimi K2.6, based on those that have been capable of take a peek. (We’ve but to check this ourselves). K2.6 is a roughly one-trillion-parameter mixture-of-experts mannequin Moonshot launched on April 20. Combination of consultants is an structure that retains solely a fraction of its parameters lively at any second; K2.6 prompts round 32 billion per token whereas carrying a 256K-token context window. (For reference, tokens are the smallest quantity of knowledge an AI can course of whereas parameters are all of the numerical values that retailer all of the data and traits a mannequin has.)

That context window issues: It means the agent can maintain an infinite quantity of knowledge in thoughts throughout an extended, multi-step workflow with out forgetting what it began on.

When you’ve got doubts about Kimi’s high quality as an AI mannequin, think about that this was the bottom utilized by the favored AI code editor Cursor to effective tune its personal specialised coding massive language mannequin, “Composer 2.”

The misperception value clearing up first

“Native” in Kimi Work refers to the place the actions occur—your machine—not essentially the place the AI mannequin runs. The K2.6 mannequin inference can nonetheless route by Moonshot’s API within the cloud, even whereas the file reads, browser clicks, and Python executions occur regionally.

If you need full on-device inference, the weights can be found on Hugging Face underneath a Modified MIT License—however a trillion-parameter mannequin calls for severe {hardware} most dwelling customers do not have.

The privateness image can be extra nuanced than “native equals secure.” As a result of WebBridge drives your precise logged-in browser, it may contact your checking account, your e-mail, and your organization’s inside instruments. Researchers at UC Riverside warned in Could that AI brokers usually perform duties with out recognizing when their actions are dangerous—a habits they referred to as “blind goal-directedness.”

Moonshot contains an “ask earlier than performing” mode that requires your approval earlier than any file is modified or code is run. That is the appropriate default to depart on, but it surely’s nonetheless removed from being 100% secure.

The brewing warfare within the agentic AI period

The desktop agent race has gotten crowded quick. Anthropic’s Claude has supplied full-desktop pc use since late 2024. OpenAI shipped Codex Background Laptop Use for macOS in April 2026, operating brokers in parallel desktop periods. Google’s Gemini pc use—descended from Undertaking Mariner—focuses on browser workflows. Microsoft’s Copilot Studio added computer-use in Could 2026, geared toward enterprise automation and utilizing each OpenAI and Anthropic fashions underneath the hood.

But it surely looks like as an alternative of being constrained to a single supplier, customers need some flexibility. That is the place instruments like OpenClaw, Hermes, or NanoClaw enter within the recreation. They’re mainly native platforms that configure your AI brokers utilizing any LLM over API.

The distinction with Kimi Work is the local-first design paired with a 300-agent swarm. Most rivals both run all the things in a cloud sandbox—which might’t contact actual logged-in periods—or provide desktop management with out coordinated parallel brokers. Kimi Work does each. The tradeoff is that this: When your laptop computer is closed, duties cease. Moonshot’s personal cloud product, Kimi Claw, runs 24/7 with out your machine.

The app is a free obtain. Significant agent options require a paid plan. Moonshot’s Moderato tier begins at $19/month and contains K2.6, Deep Analysis, and Kimi Code entry.

The Agent Swarm with restricted sub-agents unlocks at Allegretto ($39/month), whereas the total 300-agent swarm and highest-volume skilled workflows require the Allegro ($99/month) or Vivace ($199/month) tiers for many who actually want 300 brokers.

Downloads for macOS (Apple Silicon) and Home windows are stay at kimi.com, with the interior testing part which means some options should shift earlier than a wider rollout.