← Back to feed

Moonshot AI has launched Kimi Work, a desktop AI agent for macOS and Windows that runs locally on your machine. The Beijing-based lab opened it for internal testing in June 2026. Unlike cloud-based coding agents, Kimi Work mounts your local folders, drives your real browser, and executes Python — without any of it leaving your device unless you ask it to.

The agent runs on Kimi K2.6, the open-weight mixture-of-experts model Moonshot released in April under a Modified MIT license. K2.6 carries 384 experts (~32B active per token), a 256K-token context window, and native text, image, and video input.

The Architecture Difference

Most production coding agents — Claude Code, OpenAI Codex, Cursor Composer — run in cloud infrastructure. They operate on uploaded or sandboxed files, use hosted virtual browsers, and own the execution environment. The advantage is zero setup. The tradeoff is that every file, session, and credential passes through a vendor.

Kimi Work inverts this. The agent:

• Mounts your local folders and reads files directly, keeping originals intact unless you approve changes
• Drives your real browser via WebBridge, meaning it can operate on logged-in sessions that a sandboxed agent cannot touch
• Runs a built-in cron engine for scheduled background tasks
• Executes Python locally, with an “Ask before acting” gate before any file write (a YOLO mode is available for power users)

The net effect: Kimi Work can do things a cloud agent structurally cannot — interact with pages behind authentication, operate on local databases, or process proprietary documents that can’t leave the machine.

Agent Swarm: 300 in Parallel

The headline capability is scale. Kimi Work’s Agent Swarm splits large tasks into sub-tasks and distributes them across up to 300 parallel sub-agents, coordinating across as many as 4,000 steps. A sidebar task map shows each agent’s real-time status — what it is reading, which line of code it is fixing — so the user can monitor or intervene at any point.

The swarm numbers represent a generation jump from Kimi K2.5, which maxed out at 100 sub-agents over 1,500 coordinated steps. The improvement comes from PARL (Parallel-Agent Reinforcement Learning), a technique Moonshot uses to train the orchestrator to allocate agents productively rather than spawn them redundantly.

For context: Claude Code’s Dynamic Workflows reaches 1,000 sub-agents, but lives in the terminal and requires writing workflows. Kimi Work caps at 300 but wraps the swarm in a desktop GUI with no configuration step.

What It Delivers

Kimi Work is positioned at knowledge workers rather than software engineers specifically. It produces documents, spreadsheets, PDFs, and slide decks by reasoning over large, dense inputs. Feed it a research question and it browses sources, synthesizes findings, and returns a structured report. It can also build and render working websites, generating code and showing a live preview.

Moonshot describes it as a “system-level digital employee” — a phrase that draws a deliberate contrast with chat assistants, positioning the product as a persistent agent that handles complete workflows rather than single-shot questions.

Release Status

Kimi Work is in staged rollout starting with internal testing, consistent with how Moonshot has launched previous agentic features. Pricing and exact GA timeline have not been disclosed.

The underlying K2.6 model is public and open-weight; developers can self-host it independently of the Kimi Work desktop application.

Numbers

• Sub-agents: up to 300 in parallel
• Coordinated steps: up to 4,000 per task
• Model: Kimi K2.6, ~1T parameters, ~32B active, 256K context
• Platforms: macOS (Apple silicon), Windows
• License: K2.6 under Modified MIT; Kimi Work product terms TBD