Moonshot AI, the Beijing-based lab behind the Kimi chatbot, released Kimi Work on July 20, 2026, a downloadable desktop application that connects directly to a user’s local files and automates tasks inside a web browser. The app ships for both Windows and macOS, putting a Chinese lab’s autonomous software directly onto Western desktop operating systems rather than confining it to a browser tab. That placement is deliberate: Moonshot is competing on where the agent lives, not only on what model powers it.
That positioning matters against the two labs setting the pace in the West. OpenAI’s ChatGPT Agent and Anthropic’s Claude cowork both operate primarily through hosted browser sessions rather than a local install. Kimi Work instead mounts folders on the user’s own machine and runs Python scripts there, an approach closer to Manus, the Chinese startup that built its product around persistent, file-aware task execution rather than chat.
The centerpiece is WebBridge, a module Moonshot says lets Kimi browse the open web the way a person would: opening tabs, pulling data out of pages, and finishing multi-step tasks from a single prompt. The company’s product page describes a user handing over a goal, then the software clicking, scrolling and researching without further input. Moonshot has not published an independent success rate for how often those unsupervised browsing sessions finish correctly.
Kimi Work also ships a built-in Cron scheduler for recurring jobs, such as an early-morning agent call that drafts a daily briefing or an overnight Python script that processes a large dataset. A “Keep Computer Awake” toggle keeps those jobs running while the machine would otherwise sleep. Moonshot calls the underlying orchestration Agent Swarm: multiple specialized agents split a complex task and work it in parallel, then the output converts into a PowerPoint deck or an Excel file.
Moonshot is also courting financial analysts specifically. Kimi Work comes pre-loaded with data feeds covering A-shares, Hong Kong equities and US stocks, letting a user pull an earnings report or reconcile a spreadsheet through conversation instead of building an API integration. The company’s own marketing frames this as a stand-in for a desktop analyst, a claim that has not been tested against a licensed data terminal in any published comparison.
The company’s FAQ addresses the obvious risk of a file-mounted, script-running agent: unsupervised changes to a user’s system. Moonshot says an “Ask before acting” safeguard requires explicit authorization before Kimi modifies, overwrites or executes code inside local directories. The product page does not disclose what share of actions actually trigger that prompt, nor how the safeguard behaves during an unattended overnight Cron job when no one is present to answer it.
That gap is the one worth tracking. Agent products that browse the web autonomously have a history of compounding small errors across each click and extraction, and a scheduler built to run unmonitored overnight raises the cost of a single bad decision made without a human in the loop. Moonshot has not released a benchmark measuring WebBridge’s task completion rate, so the claim currently rests on the company’s own product copy rather than a third-party evaluation.
Kimi Work arrives as desktop-level agent access becomes its own competitive front, separate from whichever model sits underneath it. Anthropic, OpenAI and Manus are each building toward software that operates a full computer rather than answering questions in a chat window. A Chinese lab shipping that capability straight to Windows and Mac, rather than gating it behind an enterprise pilot, signals Moonshot intends to compete on distribution as much as on raw model capability.
For operators evaluating a desktop agent, the near-term test is not whether Kimi Work can finish a task once under supervision. It is whether the Ask-before-acting safeguard holds up across an unattended overnight Cron run before any team trusts it with real file systems and production spreadsheets.
Based on Kimi’s official product page (Moonshot AI), describing the Kimi Work desktop agent launched July 20, 2026.
