What Is Kimi Work? The Desktop AI Agent Explained
Kimi Work is Moonshot AI’s desktop agent for knowledge work. Give it access to a folder and a goal, and it can read and organize local files, research in your signed-in browser, run Python or shell commands, and produce finished spreadsheets, documents, slides, or websites. It can also keep a longer objective moving in Goal Mode or repeat a job on a schedule.
That puts it in the same emerging category as Claude Cowork: an AI agent that acts on the computer instead of stopping at an answer in a chat box. Kimi Work’s distinctive pitch is the combination of local execution, browser control, scheduled tasks, and an Agent Swarm that can split a large job across many sub-agents.
Moonshot launched Kimi Work on June 3, 2026. It is still labeled beta, so the feature set is moving quickly and Moonshot explicitly says execution stability and delivery quality remain active areas of improvement.
Kimi Work at a glance
| Kimi Work | |
|---|---|
| What it is | A general-purpose desktop AI agent |
| Company | Moonshot AI |
| Platforms | macOS and Windows |
| Core jobs | Local file work, browser automation, research, data analysis, coding, reports, slides, and spreadsheets |
| Automation | One-time tasks, Goal Mode, and scheduled tasks |
| Starting price | Free with limited credits; paid plans start at $19/month |
| Current status | Beta |
The app has separate Chat and Work modes. Chat behaves like the familiar Kimi assistant. Work is the agent workspace: you create a project, optionally attach a local folder, choose a permission level, and describe the outcome you want.
What Kimi Work can actually do
Kimi Work brings several capabilities that normally live in separate AI products into one desktop app.
Read and write local files
Mount a folder and the agent can work directly inside it. It can search PDFs, organize files, reconcile spreadsheets, generate a report, and save the result back to the same workspace. This avoids the upload-download-rename loop common to browser chatbots.
A practical prompt could be:
Read the customer interview transcripts in this folder, group the recurring complaints, quantify how often each theme appears, and save a summary spreadsheet and a five-slide briefing here. Do not modify the source files.
The final sentence matters. Kimi can modify files and execute commands, so prompts should define which files are read-only, where outputs belong, and what counts as complete.
Operate a signed-in browser with WebBridge
Kimi WebBridge connects the agent to a browser extension through a local service. It can open pages, click elements, fill forms, take screenshots, extract tables, download files, and reuse the login sessions already in your browser.
Moonshot says WebBridge actions and page content execute locally rather than being sent through a remote browser. That is narrower than saying Kimi Work is fully offline, however: Moonshot also describes it as a local agent powered by cloud-based models. Treat “local” as a description of where files, code, and browser actions are handled—not proof that every part of model inference happens on your device.
Keep long jobs moving with Goal Mode
For work that will not finish in one pass, Goal Mode stores the objective, constraints, and acceptance criteria. At the end of each round, Kimi decides whether the goal is complete, blocked, paused, or ready for another round. You can review progress, supply missing information, or redirect the task without restating the entire brief.
This is useful for multi-stage jobs such as building a market map, processing a large archive, or iterating on a website. It is not unlimited autonomy: Moonshot recommends explicit caps on rounds, time, or token use, and the computer must remain on while a goal runs.
Run recurring jobs on a schedule
Kimi Work includes a Cron-based scheduler for one-time, daily, weekly, and monthly tasks. You can configure the schedule manually or ask Kimi to turn a chat instruction into one. Examples include:
- Produce a morning briefing from selected sites and local files
- Reconcile a downloaded report against a spreadsheet every Friday
- Run a Python data-cleaning script overnight
- Refresh a dashboard or market summary each morning
Scheduled tasks run on the local computer. If it is asleep, the agent cannot quietly move the job to Moonshot’s cloud; Moonshot’s own guidance is to enable Keep Computer Awake for overnight work.
Split large projects across an Agent Swarm
Kimi Work can divide a large request among specialized sub-agents and combine their results. Moonshot’s current documentation advertises up to 300 sub-agents and more than 4,000 parallel tool calls for its highest-scale swarm workflows.
The useful part is parallel decomposition, not the headline number. A research project can assign different companies, markets, or source sets to separate agents, then have another pass normalize and synthesize the findings. Swarm use is credit-intensive and paid plans limit both uses and concurrent subtasks.
Create dashboards and install plugins
Kimi Work’s newer Dashboard feature turns natural-language requests into persistent interactive widgets. You can create a custom countdown, task panel, experiment tracker, or market overview, refine it by annotating the rendered result, then keep it in the dashboard or pin it to the desktop.
Plugins extend the agent to third-party tools and professional data. Moonshot names services such as Notion, Canva, and WPS, along with financial, academic, legal, and economic databases. Some plugins use OAuth, MCP, or both; if you are unfamiliar with that connection layer, our MCP guide explains it without the protocol jargon.
Kimi Work pricing
Kimi Work is available on the free Kimi plan. New accounts receive limited trial credits, and free users can create up to two scheduled tasks. Current Kimi membership pricing is:
| Plan | Monthly price | Approx. agent tasks | Agent Swarm |
|---|---|---|---|
| Adagio | Free | 6 | Not included |
| Moderato | $19 | 60 | 25 uses |
| Allegretto | $39 | 150 | 50 uses |
| Allegro | $99 | 360 | 120 uses |
| Vivace | $199 | 720 | 240 uses |
Those task counts are estimates, not hard task allowances. Kimi deducts credits based on the tokens and complexity a job consumes, so a long research-and-deck workflow can cost much more than a short file operation. Credits refresh monthly and do not roll over.
Kimi’s membership documentation is also in transition: the help center says a new plan structure is coming and some credit pools will be separated. Check the live pricing page before buying rather than assuming today’s allowances will remain unchanged.
The limits to understand before granting access
Kimi Work is unusually capable for a free desktop agent, but five constraints shape where it fits.
- It is a beta with broad permissions. Moonshot itself lists stability, task decomposition, browser operation, local file handling, and long-form delivery as areas still being tested. Start with a contained workspace and review outputs before pointing it at business-critical folders.
- “Local” does not mean “offline model.” File operations, code, and WebBridge run on the computer, while Kimi’s own material describes the agent as powered by cloud-based models. Organizations with sensitive data should verify the applicable privacy and retention terms rather than relying on the local-agent label.
- Scheduled work depends on the device. The computer must stay awake and the Kimi app must be able to run. This is different from a cloud agent that continues when your laptop is closed.
- Usage is metered. Complex Agent and Swarm jobs consume shared credits. The free tier is best for evaluation, not high-volume recurring operations.
- Browser automation inherits browser risk. A logged-in browser lets the agent do useful work, but it also gives web pages a path to influence the agent. Use Request permission, restrict the sites and folder scope, and avoid unattended financial, publishing, or destructive actions until the workflow is proven.
System support is broader than some competitors but not universal. Windows requires Windows 10 or later. Mac requires macOS 12 or later and Apple silicon, so Intel Macs are excluded.
Kimi Work vs Claude Cowork vs Carly
All three products can reduce manual work, but they start from different places.
| Kimi Work | Claude Cowork | Carly | |
|---|---|---|---|
| Primary environment | Your desktop, folders, and browser | Claude’s agent workspace, connectors, and files | Cloud apps, inboxes, calendars, and event-driven workflows |
| Best trigger | A prompt, goal, or local schedule | A prompt or scheduled task | An app event, inbound message, or schedule |
| Local file access | Direct folder access | Folder access in supported Cowork environments | No desktop-folder control |
| Browser control | WebBridge in your signed-in browser | Browser automation | Usually works through APIs instead of clicking pages |
| Runs with laptop closed | No | Yes for remote tasks | Yes |
| Distinctive strength | Local files + browser + multi-agent research | Strong supervised document and browser work | Always-on cross-app operations and a reachable agent inbox |
Choose Kimi Work when the job lives in folders on your computer and needs research, code, or browser actions to become a finished artifact. Choose Cowork when you want a polished interactive agent inside Claude and prefer its model and connector ecosystem. Choose Carly when the job should begin because something happened—an email arrived, a meeting was booked, a form was submitted—and continue across business apps without depending on a desktop session.
Carly gives each agent its own email address and runs in the cloud, so clients or coworkers can hand it work directly. It connects to a broad native integration catalog, and other services with a reachable public API can connect with your own key. Zapier-style workflows are free and unlimited; AI agents start at $35/month. Kimi Work is the stronger fit for hands-on local files. Carly is the stronger fit for ongoing, event-driven operations.
Who Kimi Work is best for
Kimi Work makes the most sense for:
- Analysts turning local reports and live market data into recurring briefings
- Researchers processing large sets of papers into tables, summaries, and presentations
- Consultants and operators whose deliverables are spreadsheets, reports, and decks
- Developers who want Kimi Code’s agent capabilities in a visual desktop workspace
- Anyone with repeatable jobs that cross local files and a signed-in browser
It is a weaker fit when tasks must run reliably with every device offline, when colleagues need to send work directly to the agent, or when the real workflow is triggered by changes across many SaaS tools.
Frequently asked questions
Is Kimi Work free?
Yes. The desktop app is available on Kimi’s free Adagio plan with limited credits and up to two scheduled tasks. Paid plans start at $19/month and add more agent credits, Agent Swarm uses, concurrency, and professional-database calls.
Does Kimi Work run on Mac and Windows?
Yes. It supports Windows 10 or later and Apple-silicon Macs running macOS 12 or later. Intel Macs are not supported.
Can Kimi Work run while the computer sleeps?
No. Its scheduled tasks and Goal Mode run locally, so the computer must remain awake. Kimi provides a Keep Computer Awake setting for unattended or overnight jobs.
Is Kimi Work fully local and offline?
No. It executes file operations, commands, and browser actions on the local machine, but Moonshot describes the product as using cloud-based models. It should not be treated as a fully offline local-language-model app.
What is the difference between Kimi Work and regular Kimi?
Regular Kimi is the web and mobile assistant for chat, research, and generated content. Kimi Work adds direct access to local folders, local command execution, browser automation through WebBridge, persistent Goal Mode, scheduled tasks, dashboards, and plugins in a desktop app.
More: Claude Cowork review · What is ChatGPT Work? · Best AI assistant for work · AI news, July 17
Ready to automate your busywork?
Carly schedules, researches, and briefs you—so you can focus on what matters.
See what people say
"Before Carly, I relied on a Calendly link, but the whole process felt impersonal and not very professional. Carly changed that by handling all the back-and-forth, so I'm no longer stuck in endless email threads trying to line up schedules.
Now Carly reaches out to candidates, shares my real-time availability, lets them pick a slot, then sends a Zoom link and drops it straight into my calendar. She sends reminders to both of us before each call, which has significantly reduced no-shows and last-minute confusion.
On top of scheduling, Carly acts like a full executive assistant, sending me my schedule the night before so I can prepare for each call. It reminds me of the old x.ai assistant, but Carly is noticeably smarter, faster, and better suited to my healthcare recruitment business."
