Kimi Code: what it is, its official website, and how to install it on Windows and macOS

On September 17, 2026, Moonshot AI launched the desktop version of Kimi Code, its AI coding agent. Before this, it ran in the terminal and as editor plugins; now it is also available as a standalone app for macOS and Windows with a browser and terminal built in. Commands and plans were checked against the official pages on September 19, 2026.

Ailoot documentation: API, keys, and setup.

What Is Kimi Code#

In terms of purpose, it is a direct counterpart to Claude Code: you open a project folder, describe a task in ordinary language, and the agent studies the codebase, suggests edits, and carries them out. You choose how often it asks for your permission. I discussed how they differ in practice in Kimi K3 vs Claude Code; this article covers only installation and setup.

Kimi Code uses Moonshot models: the flagship K3 with a context window of up to 1 million tokens, and the specialized K2.7 Code model, which has a high-speed mode of up to 260 tokens per second. The CLI source code is open under the MIT license, and the GitHub repository has more than 7,000 stars.

Official website, documentation, and source code:

• product website: kimi.com/code
• CLI documentation: kimi.com/code/docs and moonshotai.github.io/kimi-code
• repository: github.com/MoonshotAI/kimi-code

On September 11, 2026, Kimi Code updated its coding model; on the 17th, the desktop version and the new major CLI version 2.0 launched at the same time; and on the 18th, update 2.0.1 was released. If you have the old Python version (kimi-cli) installed, it is no longer supported: its data can be migrated to the new CLI with a single command, kimi migrate.

How Much Does Kimi Code Cost#

The app and CLI themselves are free to download, but the free Adagio plan does not include Kimi Code. At the time of checking, the subscription page kimi.com/membership listed these prices:

Kimi Code is available on all four paid plans. In September 2026, Kimi revised the subscription lineup and stopped selling Kimi Code separately, so older reviews with different plan names are out of date—check the subscription page before buying.

The subscription has two limits. All its features (chat, Deep Research, websites, Kimi Code) draw from one shared pool of credits based on the actual number of tokens used. Kimi Code also has its own limit: 5 hours per week. It does not affect the other features, but you can run out of it if you work intensively.

Instead of a subscription, you can pay for tokens as you use them: when signing in to the CLI, choose “Kimi Platform API key” from platform.kimi.ai. This is convenient if you work infrequently or want to connect the model to a third-party tool. Check the platform page for the terms and limits of this option.

Before paying, check in practice whether your card works, or use the API key option if it is available to you.

How to Install Kimi Code CLI#

The CLI is the primary and most complete form of Kimi Code. If you have already worked with Claude Code in the terminal, the workflow will be familiar. If not, the guide on how to install Claude Code can help you get acquainted: they share the same workflow.

1. Install the CLI with one command

   On macOS or Linux, run this in the terminal:

   curl -fsSL https://code.kimi.com/kimi-code/install.sh | bash
   

   On Windows, run this in PowerShell:

   irm https://code.kimi.com/kimi-code/install.ps1 | iex
   

   The installer downloads the latest release, verifies its checksum, and adds the kimi command to PATH. Node.js is not required for this option.

   On Windows, there is one additional requirement: install Git for Windows before the first launch. Kimi Code uses its Git Bash as its working shell. If Git Bash is installed in a nonstandard location, set the path to bash.exe in the KIMI_SHELL_PATH environment variable.

   There is also an npm option for those who already have Node.js 22.19.0 or later installed:

   npm install -g @moonshot-ai/kimi-code
   

2. Check the installation and launch

   In the cd your-project command, replace your-project with the path to your project folder.

   kimi --version
   cd your-project
   kimi
   

   For a single command without interactive mode, use the -p flag:

   kimi -p "Look at the project structure and explain the main directories"
   

   To continue a previous session, use the -c flag: kimi -c will return you to the same conversation.

3. Sign in to your account

   Enter this in the CLI:

   /login
   

   A choice of two options will appear. “Kimi Code (OAuth)” signs in through your Kimi account: open the link on any device, sign in, and enter the code. This uses your subscription limits. “Kimi Platform API key” is a key from platform.kimi.ai; you pay for tokens. To sign out, use /logout.

4. Give it its first task

   Start with an overview of the project: ask the agent to study the main directories and explain how everything is organized. This lets you see how it reads files and check that sign-in and model access work.

   When a new version comes out, you can update with the kimi upgrade command.

How to Install the Desktop Version#

The desktop version was released on September 17, 2026, and is for people who do not want to work exclusively in the terminal. According to Moonshot AI documentation, one window combines a conversation with the agent, a browser, and a terminal: open a local project, discuss a task, and submit changes without switching between windows. You can see every agent step: tool calls, reasoning, and modified files; sensitive operations require confirmation. You can set how often the agent asks before acting using three permission modes: Always Ask, Ask When Needed, and Never Ask. For a first try, the first or second mode is more sensible.

There are two task modes. Plan creates a plan first, which you can review and adjust before execution. Goal lets the agent work toward the objective on its own; you can send a long task to the background, pause it, or stop it. The built-in browser is in the right-hand panel and gives the agent page context: you can select an element on the page, comment on selected text or an area of the screenshot, and specify exactly what to fix.

Installers are available on the official website: for macOS, there are builds for Apple Silicon and Intel; for Windows, there is an x64 installer. Download your build from kimi.com/code, install it, open a local project folder, and sign in to the same Kimi account. Limits and billing are shared with the CLI: the same subscription and the same pool of credits.

How to Connect Kimi Code to an Editor#

The easiest way to install the VS Code extension is to get “Kimi Code” from the official marketplace: the agent will appear in the sidebar and can complete code, edit files, and search online. The Kimi download page also lists Cursor, Web, Zed, JetBrains, and Paseo.

Editors that support the Agent Client Protocol connect through the CLI. In Zed, add this to ~/.config/zed/settings.json:

{
  "agent_servers": {
    "Kimi Code CLI": {
      "type": "custom",
      "command": "kimi",
      "args": ["acp"],
      "env": {}
    }
  }
}

After this, a new conversation in Zed's Agent panel will run through Kimi Code. You only need to sign in to your account once in the CLI.

What the Agent Can Do Out of the Box#

In addition to the standard “read, edit, run tests,” the CLI has several features:

• Video input: you can drag a screen recording or demo video into the chat, and the agent will analyze it frame by frame—useful when a bug is easier to show than describe.
• MCP without manual JSON: Model Context Protocol servers can be added and configured through a conversation using /mcp-config. We discussed what MCP is and why to use it in the article about MCP servers for Claude Code; the protocol is shared.
• Subagents: the built-in coder, explore, and plan carry out tasks in an isolated context, keeping details of their work out of the main conversation.
• Lifecycle hooks: custom commands at key stages of the workflow for checking risky actions, notifications, and auditing.
• Skills and plugins: ready-made packages from the marketplace or any GitHub repository.

What to Check if Something Goes Wrong#

• kimi --version does not respond—open a new terminal window; PATH is picked up after a restart.
• On Windows, the agent exits with an error when running commands—check that Git for Windows is installed and that KIMI_SHELL_PATH is set if it is in a nonstandard location.
• Sign-in fails or tasks do not run—check /login: OAuth works only with an active paid subscription; an API key works when the platform balance is positive.
• Everything worked, but the agent stopped carrying out tasks—your 5-hour weekly Kimi Code limit or shared pool of credits may be exhausted; the remaining balance is shown in the subscription section of the Kimi website.
• The old Python CLI conflicts with the new one—remove the old version and reinstall using the official installation command.

What to Do Next#

Choose the app, CLI, or plugin and try a real task from your project: have the agent study the codebase and make a small edit. This will help you understand whether its way of working suits you. I tried Kimi K3 on real tasks myself and eventually went back to Claude Code—here I explained what did not work out. A task-by-task comparison is available in Kimi K3 vs Claude Code, and among free options, I would look at the review of Qwen Code as an alternative to Claude Code.

If you need access to neural networks through Ailoot, check out Ailoot.