Moonshot AI Launches Kimi Work, a Desktop Agent That Runs 300 Parallel Sub-Agents on Your Machine

Moonshot AI, one of China’s “AI Tiger” startups, released Kimi Work this week: a downloadable desktop agent for macOS and Windows that reads local files, controls your real browser, schedules recurring tasks, and coordinates up to 300 parallel sub-agents. The app is available as a free download at kimi.com, with paid tiers unlocking the multi-agent features.

How Kimi Work Operates

The core pitch is locality. Where most AI agent tools spin up sandboxed browser sessions in the cloud, Kimi Work runs actions on the user’s machine. It reads mounted folders, executes Python locally, and drives actual Chrome or Edge sessions through WebBridge, Moonshot’s browser extension that uses Chrome DevTools Protocol. That means agents interact with your logged-in sessions, cookies, and local files directly.

The Agent Swarm feature lets Kimi Work spawn multiple sub-agents in parallel, each handling a different slice of a task. The full 300-agent capacity is limited to higher-tier subscriptions. A built-in cron engine handles scheduled tasks on daily, hourly, or conditional triggers, with a “Keep Computer Awake” toggle for overnight runs, according to Decrypt.

Under the hood, Kimi Work runs on Kimi K2.6, a roughly one-trillion-parameter mixture-of-experts model that Moonshot released on April 20. K2.6 activates approximately 32 billion parameters per token and carries a 256K-token context window. The model is notable enough that Cursor used it as the base for fine-tuning its Composer 2 coding model.

Pricing Tiers

Moonshot structured Kimi Work across four paid plans: Moderato ($19/month) includes K2.6, Deep Research, and Kimi Code. Allegretto ($39/month) unlocks Agent Swarm with limited sub-agents. Allegro ($99/month) and Vivace ($199/month) provide the full 300-agent swarm and highest-volume professional workflows. The desktop app itself is a free download, but meaningful agent features require a subscription.

The Local Tradeoff

“Local” in Kimi Work refers to where actions execute, not necessarily where inference runs. K2.6 model inference can still route through Moonshot’s cloud API, Decrypt notes. The open weights are available on Hugging Face under a Modified MIT License, but a trillion-parameter model demands hardware most users lack.

The privacy picture is also more complicated than “local equals safe.” Because WebBridge drives a real logged-in browser, it can access bank accounts, email, and internal company tools. Moonshot includes an “ask before acting” mode that requires user approval before files are modified or code is executed.

When the laptop is closed, tasks stop. Moonshot’s cloud product, Kimi Claw, runs 24/7 without the user’s machine.

The Desktop Agent Landscape

Kimi Work enters a crowded race. Anthropic has offered full-desktop computer use since late 2024. OpenAI shipped Codex Background Computer Use for macOS in April 2026. Google’s Gemini computer use descends from Project Mariner. Microsoft’s Copilot Studio added computer use in May 2026. OpenClaw, Hermes, and NanoClaw offer local platforms that configure agents across any LLM via API.

The distinction Kimi Work draws is pairing local-first execution with coordinated parallel agent swarms. Most competitors either run agents in cloud sandboxes that cannot access real logged-in sessions, or offer desktop control without parallel agent coordination. Kimi Work attempts both, but trades always-on availability for locality.