Moonshot AI launches Kimi Work: a local desktop agent that can spin up to 300 AI sub-agents
Beijing startup Moonshot AI this week released Kimi Work, a downloadable desktop agent for macOS and Windows that lives on your machine, reads local files, drives your browser, and runs scheduled jobs. The app is available for free download now (currently in internal testing) and builds on Moonshot’s recent WebBridge extension — which already allowed local control of real Chrome and Edge sessions — by packaging the functionality into a full desktop product.
What Kimi Work does
- Local-first automation: Unlike cloud-only assistants that run in sandboxed servers, Kimi Work runs on your PC or Mac and can interact with local files, folders you mount, and your real browser sessions. It can manipulate PDFs, organize your desktop, pull market data from open tabs, compile HTML reports and email them, run Python scripts in the background, and export finished research straight to PowerPoint or Excel.
- WebBridge integration: The agent controls your real browser via Chrome DevTools Protocol, so your logged-in sessions and cookies remain on your machine while the agent performs actions.
- Agent Swarm: Kimi Work can spawn many sub-agents in parallel — up to 300 — letting the app split large tasks into concurrent slices for faster, coordinated work.
- Scheduling & background work: A built-in Cron engine supports daily, hourly, or conditional triggers, with a “Keep Computer Awake” toggle for long overnight runs.
- Market data out of the box: Native feeds for A-shares, Hong Kong stocks, and U.S. equities are pre-integrated — no API keys required.
- Local file and compute layer: Agents can read mounted folders and execute Python locally.
Under the hood: K2.6 model and where inference runs
Insiders report Kimi Work runs on Moonshot’s K2.6 model. K2.6 is a mixture-of-experts model Moonshot released on April 20 with an approximate one‑trillion‑parameter architecture. In practice it activates roughly 32 billion parameters per token and supports a 256K-token context window — enabling the agent to hold long, multi-step workflows in memory. (For context: tokens are the smallest units an LLM processes; parameters are the model’s learned weights.)
“Local” refers to where actions occur (your machine), not necessarily where model inference happens. K2.6 inference can be routed through Moonshot’s cloud API even while file reads, browser clicks, and Python run locally. Full on-device inference is possible — the weights are available on Hugging Face under a Modified MIT License — but a trillion‑parameter model requires serious hardware that most consumer laptops don’t have.
Security and privacy — not as simple as “local = safe”
The combination of real browser control and local file access is powerful but also raises risks. Because WebBridge drives your logged-in browser, an agent could touch bank accounts, email, or company tools. Academic researchers (UC Riverside) warned in May about agents acting without recognizing risky actions, a behavior dubbed “blind goal-directedness.” Moonshot includes an “ask before acting” mode that requires user approval before file modifications or code execution — a recommended default — but it's not a complete safety guarantee.
How Kimi Work fits into the desktop agent race
Desktop agents have proliferated: Anthropic’s Claude offered desktop computer use since late 2024; OpenAI shipped Codex Background Computer Use for macOS in April 2026; Google’s Gemini features browser-focused computer use from Project Mariner; Microsoft added computer-use capabilities in Copilot Studio in May 2026. There are also local platforms like OpenClaw, Hermes, and NanoClaw that let users configure agents around any LLM API.
Kimi Work differentiates itself by combining local-first access to real sessions with a coordinated, high‑parallelism Agent Swarm (up to 300 sub-agents). Many competitors either keep everything sandboxed in the cloud (and thus can’t touch logged-in sessions) or provide desktop control without a coordinated parallel-agent architecture. The obvious trade-off: when your laptop is closed, local tasks stop — whereas Moonshot’s cloud product, Kimi Claw, can run 24/7.
Pricing and availability
Kimi Work is a free download for macOS (Apple Silicon) and Windows at kimi.com, but core agent capabilities require a paid plan:
- Moderato — $19/month: includes K2.6, Deep Research, and Kimi Code access.
- Allegretto — $39/month: unlocks Agent Swarm with a limited number of sub-agents.
- Allegro — $99/month and Vivace — $199/month: unlock the full 300-agent swarm and higher-volume professional workflows.
The app is in internal testing, so features and polish may change before a wider rollout.
Bottom line
Kimi Work is an ambitious local desktop agent that blends real-browser control, deep local file access, scheduled background work, and extreme parallelism via a 300-agent swarm — all driven by Moonshot’s large-scale K2.6 model. That combination is appealing for power users and researchers (and potentially traders who need integrated market data and fast workflows), but it also sharpens the privacy and safety trade-offs inherent to giving agents control of real sessions and local files.
Read more AI-generated news on: undefined/news
