Kimi Work: a 300-Agent Swarm on Your Desktop, No Sandbox

Moonshot AI opened testing of Kimi Work this month, and the positioning is blunt: a general local agent for knowledge workers. Not a cloud sandbox you upload files to. It lives on your machine (macOS Apple Silicon and Windows), touches your real files, executes code locally, and drives the browser you are already logged into via Kimi WebBridge, cookies and sessions included. Under the hood is K2.6, Moonshot's trillion-parameter MoE agent model, with an Agent Swarm that scales to 300 parallel sub-agents across 4,000 coordinated steps, running from your desktop. It also ships native market data for A-shares, Hong Kong and US equities, so earnings pulls and spreadsheet reconciliation work out of the box.

The AI desktop race is getting crowded fast: Devin Desktop, Claude Cowork, Perplexity's Personal Computer, now Kimi. But Moonshot made the opposite bet from everyone else. This month's industry consensus is isolate everything: Microsoft MXC puts agents in micro-VMs, Apple just shipped Container machine, Claude Cowork runs inside a sandboxed VM. Kimi Work says the access IS the value. Your real browser with your real sessions is exactly what makes it useful, and exactly what security researchers will lose sleep over, the same week a one-cent bank transfer was shown hijacking a banking agent.

That tension, capability from access versus safety from isolation, is becoming THE design split in desktop agents. Kimi just planted its flag on the access side harder than anyone.

https://www.kimi.com