Moonshot AI’s Kimi Work is a downloadable desktop agent for macOS and Windows that runs up to 300 sub-agents in parallel on your local machine, automating file workflows, browser tasks, and scheduled jobs powered by the K2.6 model.
The desktop agent race moved out of the cloud. Moonshot AI, the Beijing-based lab behind the Kimi series of open-weight models, released Kimi Work in June 2026 as a native application that lives on your machine. It reads local files, drives your real browser through a Chrome extension called WebBridge, runs scheduled tasks, and orchestrates hundreds of sub-agents working toward a single goal. It is the first major attempt to bring that kind of parallel agent swarm to a consumer desktop product.
Key Takeaways:
- Kimi Work is a downloadable desktop agent for macOS and Windows, not a web chat.
- Powered by Kimi K2.6, a ~1 trillion parameter mixture-of-experts model with 32B active parameters and 256K context window.
- Up to 300 sub-agents run in parallel, coordinating workflows of up to 4,000 steps.
- WebBridge extension gives the agent control of your real signed-in browser sessions.
- Built-in cron scheduler, Goal Mode, and local file access set it apart from cloud-only agents.
- Pricing starts free, with paid tiers from $19 to $199/month for deeper agent capabilities.
What Kimi K2.6 powers
Kimi Work runs on Kimi K2.6, an open-weight mixture-of-experts model that Moonshot AI released on April 20, 2026. The model has roughly one trillion parameters total, but activates about 32 billion per token. It carries a 256,000-token context window, which matters for multi-step workflows that need to hold dozens of documents, codebases, or research threads in memory simultaneously. According to Decrypt, K2.6 edged out GPT-5.4 and Claude Opus 4.6 in certain coding and reasoning benchmarks, though Moonshot does not publish a comprehensive leaderboard.
The weights are available on Hugging Face under a modified MIT-style license, so anyone can run K2.6 on their own hardware if they have the GPU budget. Moonshot’s own research division published a technical blog post detailing the model’s coding capabilities and agent swarm architecture. In practice, most Kimi Work users route inference through Moonshot’s cloud API rather than running the model locally.
The 300-agent swarm mechanic
The headline feature is the agent swarm. Give Kimi Work a single goal: pull Q2 earnings data for ten tickers, compare them to last year, and draft a PowerPoint deck. The app splits the work into sub-tasks assigned to parallel agents. According to MarkTechPost’s coverage of the launch, the swarm can coordinate up to 4,000 steps across those agents, up from the 1,500-step ceiling in earlier generations. Each agent handles a different slice: one queries a database, another scrapes a webpage, a third formats a slide. Results are assembled into a single coherent output.
The output is not limited to text. Kimi Work can deliver code files, PDFs, PowerPoint presentations, Excel spreadsheets, Word documents, or even static websites. According to Moonshot’s official Kimi Work introduction page, the agent swarm assigns sub-tasks, verifies results, and feeds them back into the next round of work in a continuous loop that remains observable and open to human intervention.
WebBridge: real browser, real sessions
WebBridge is a Chrome extension that gives Kimi Work access to your actual browser. Unlike cloud agents that operate in hosted virtual browsers, WebBridge lets the agent use your real signed-in sessions. It searches, scrolls, extracts data, fills forms, and downloads files across your open tabs. Cookies and login state stay local.
This capability is Kimi Work’s strongest differentiator and its biggest security concern. Because the agent acts inside your real browser, it can access bank accounts, email inboxes, and internal company portals. Decrypt reported that Moonshot includes an “ask before acting” mode that requires user approval before the agent modifies files or runs code, but the default “full access” setting lets it operate without interruption. According to CryptoBriefing, this design choice positions Kimi Work firmly in the local-first camp, where data stays on your device rather than being uploaded to a vendor’s sandbox.
Goal Mode and scheduled tasks
Goal Mode shifts Kimi Work from a reactive tool to an autonomous worker. You specify an objective, acceptance criteria, and constraints, then send it off. It runs tasks, gathers information, modifies code, and produces reports while you sleep or attend meetings. When you return, the work is done and you review the deliverables.
The built-in cron scheduler handles recurring jobs. It accepts standard five-field cron expressions and pairs them with plain-language task descriptions. A daily briefing job might read: scheduled at 0 7 * * * to draft a markdown briefing and save it to ~/KimiWorkspace/briefing.md, asking for approval before writing. A “Keep Computer Awake” toggle ensures overnight jobs do not stall. According to MarkTechPost, triggers can include LLM agent calls, Python scripts, or shell commands, making the scheduler flexible for technical and non-technical users alike.
Finance and academic data access
Kimi Work ships with pre-integrated data feeds that remove the need for custom API setup. According to Moonshot’s documentation, the built-in coverage spans global equities, futures, and indices for market and securities work. It extends to World Bank macroeconomic data: national GDP, population, employment, and trade figures. It reaches into academic databases with journals, papers, preprints, dissertations, and patents. For Chinese markets, it includes A-share and Hong Kong stock data natively.
This means a single request can move from raw numbers to a finished, data-driven analysis without leaving the app. A research assistant could pull S&P 500 fundamentals, cross-reference them with academic papers on valuation models, and output an Excel workbook with charts. The official Kimi Work feature page describes this as converting “a single request” into “finished, data-driven analysis without ever leaving the app.”
How Kimi Work compares to cloud agents
The core distinction between Kimi Work and products like Anthropic‘s Claude Code or OpenAI‘s Codex is where the agent runs and what it can reach. Kimi Work operates on your desktop, mounts your local folders, and uses your real browser. Cloud agents typically run on vendor servers with sandboxed files and hosted virtual browsers. Neither approach wins outright.
Local execution keeps data on your device and reaches real files, which matters for workflows involving sensitive documents or internal systems. Cloud execution trades that control for zero-setup convenience and managed safety nets. According to Decrypt, Kimi Work’s local-first model means the “local” label refers to where actions happen, not necessarily where inference runs. The K2.6 model can still route through Moonshot’s cloud API. If you want full on-device inference, the open weights are available on Hugging Face, but that requires serious hardware.
The competitive landscape shifted quickly. Anthropic‘s Claude Code, OpenAI‘s background computer-use agent, and AgentBox’s parallel coding sandboxes all target similar workflows. Kimi Work distinguishes itself by combining the agent swarm, browser control, and local file access in a single downloadable app rather than a terminal tool or web dashboard.
Pricing and availability
Kimi Work is a free download for macOS (Apple Silicon) and Windows, available at kimi.com. Meaningful agent features, however, sit behind paid tiers. According to Decrypt’s reporting on the launch, the pricing structure is:
- Moderato: $19/month, K2.6 access, Deep Research, and Kimi Code.
- Allegretto: $39/month, Agent Swarm with limited sub-agents.
- Allegro: $99/month, Full 300-agent swarm for highest-volume workflows.
- Vivace: $199/month, Premium tier for power users.
The product launched in an internal testing phase around June 9-10, 2026. Moonshot AI, which closed a $2 billion funding round at a $20 billion valuation in May 2026, is positioning Kimi Work as the bridge between its web-based Kimi product and the growing desktop agent market.
Who should care
Kimi Work targets knowledge workers whose bottleneck is access to files and live sessions: developers managing local codebases, analysts pulling data from multiple sources, researchers compiling documents. It is not a replacement for a chatbot that answers quick questions. If you need an agent that reads your local folders, operates your real browser, runs code in the background, and returns finished deliverables instead of advice, it deserves a look. Compare notes with how Claude Code approaches the same workspace problem, or see how OpenClaw-style frameworks handle distributed agent orchestration.
The security tradeoffs are real. Giving an AI agent control of your browser and local files means trusting it not to make mistakes you cannot undo. Moonshot’s approval-gate model mitigates this, but the default full-access setting puts the responsibility on the user. According to CryptoBriefing, this is a hallmark of the local-first agent movement: you keep your data, but you also keep the risk.
Looking ahead
Kimi Work arrived at a moment when the desktop agent category is filling fast. Moonshot’s combination of a massive open-weight model, a 300-agent swarm, and real-browser automation is ambitious. Whether it scales to everyday use depends on reliability, security, and how well the product evolves beyond its current internal testing phase. The $20 billion valuation suggests Moonshot believes it has something worth betting on.
The broader implication is that AI agents are moving out of chat windows and into your actual workflow. If Kimi Work’s approach gains traction, the next wave of desktop productivity tools may not be apps you open manually. They may be agents you set running and come back to later.
Frequently Asked Questions
What is Kimi Work?
Kimi Work is a downloadable desktop AI agent for macOS and Windows built by Moonshot AI. It runs locally on your machine, reads your files, operates your real browser through WebBridge, and can coordinate up to 300 sub-agents in parallel to complete complex knowledge-work tasks.
What model powers Kimi Work?
Kimi Work is reportedly powered by Kimi K2.6, Moonshot AI’s open-weight mixture-of-experts model with approximately one trillion total parameters, 32 billion active parameters per token, and a 256,000-token context window. The model was released on April 20, 2026.
Is Kimi Work free to use?
The Kimi Work desktop app is free to download, but advanced agent features require a paid subscription. Pricing tiers range from $19/month (Moderato) to $199/month (Vivace), with the full 300-agent swarm unlocked at the $99/month Allegro tier or above.
How is Kimi Work different from cloud AI agents?
Unlike cloud agents that run on vendor servers with sandboxed environments, Kimi Work operates on your local machine. It has direct access to your files, runs Python and shell commands locally, and controls your real signed-in browser through WebBridge. This keeps data on your device but also places security responsibility on the user.
Does Kimi Work use cloud or local inference?
In most cases, Kimi Work routes model inference through Moonshot’s cloud API. The K2.6 weights are open-source on Hugging Face, so technically anyone can run them locally, but doing so requires substantial GPU hardware. Moonshot has not confirmed whether all Kimi Work traffic goes through its servers or if local inference is an option for users with capable machines.
Read the full article at https://tekmag.thsite.top/kimi-work-moonshot-ais-desktop-agent-that-runs-300-ai-sub-agents-on-your-desktop/.
Top comments (0)