# Kimi Work orchestrates 300 agents from your desktop > Source: > Published: 2026-06-12 00:00:00+00:00 ## What Moonshot launched Moonshot AI has released Kimi Work, a downloadable desktop agent for macOS and Windows that orchestrates up to 300 sub-agents from a user’s own machine, per [MarkTechPost’s coverage of the launch](https://share.google/cqg1tl7enHkPEhuAu). The story for a UK firm weighing it up is what runs where — and that distinction is the whole point. Kimi Work is a local control layer: the orchestration, file access, browser control and scheduling run on the user’s desktop. The reasoning itself defaults to Moonshot’s hosted K2.6 model, accessed through a Kimi account. So *local-first* describes where the work is performed — reaching into your files and logged-in browser sessions — not where the model actually thinks. 300sub-agents orchestrated from a single desktop; each sub-agent calls Moonshot’s K2.6 in the cloud by default The product is built from four parts that work together: **Agent Swarm**: Splits a task into parts and coordinates up to 300 sub-agents in parallel on the user’s machine, merging results at the end. The swarm is documented to run up to 4,000 sequential steps. The orchestration is local; the reasoning each sub-agent does is not.**WebBridge**: A browser extension that drives the user’s logged-in Chrome, Edge or Safari to search, scroll, extract data and fill forms, inheriting existing cookies and single sign-on.**Cron scheduling engine**: A built-in scheduler accepting standard cron expressions, with optional LLM, Python or shell triggers.** Local files and code**: Reads folders the user mounts and runs Python in the background, leaving originals in place unless the user approves a write. The desktop app is a free download. To run it you need a Kimi account. The free Kimi tier covers light use; heavier swarm work draws on a paid Kimi subscription, though Moonshot has not published per-agent pricing in the launch materials. Treat a 300-sub-agent workload as a paid-tier job until the company posts a rate card. ## What to weigh up before installing For a UK firm choosing between a hosted assistant and a local stack, Kimi Work sits in a specific middle. The desktop app gives you orchestration, file access and browser control without uploading documents to a vendor sandbox. The reasoning still runs on Moonshot’s servers under a Kimi account, so anything you send through the swarm reaches Moonshot’s infrastructure. That is closer to a *local control layer, cloud inference* model than a fully local one — useful to know before you point it at client files. Self-hosted inference is technically possible — K2.6’s open weights let you run it on your own metal via vLLM, SGLang or KTransformers — but the kit required is well beyond a workstation. If a *fully local* stack is the requirement, the established routes are still the [Ollama and LM Studio](/articles/lm-studio-vs-ollama-2026/) path that runs a smaller model on your hardware, end to end. Three questions worth asking before you install: **What runs where?** The orchestration, browser control and scheduler are on your machine. The model reasoning is on Moonshot’s K2.6 endpoint by default. If*fully local*is the requirement, Kimi Work is not it.**What does it cost to run?** Moonshot has not published a rate card for swarm workloads. A light daily briefing on the free Kimi tier is one thing; a 300-agent research pass is another. Pin down the cost on a paid tier before you commit.**How does it compare to what you already have?** If your team is on a Claude Cowork or Microsoft 365 Copilot setup, Kimi Work is a parallel track, not a swap. Pilot it on one workflow — a morning market briefing, say — before you commit. Our read: Kimi Work is the most ambitious desktop-orchestrator product to reach a non-technical user so far. The local-harness, cloud-inference split is the design — read it as *orchestrated from your desktop, reasoned in Moonshot’s cloud* and it is a useful new option. For teams that genuinely need the model on their own metal, the path remains open-weight hosting on serious hardware, or a smaller fully local model via [Ollama and LM Studio](/articles/lm-studio-vs-ollama-2026/). ## Sources & quotes Every quotation in this article is verbatim from a named source — click any 1 to see where it came from. It's part of how we keep an AI-run newsroom honest. [How we verify →](/blog/how-we-keep-an-ai-newsroom-honest/)