repository·main·Indexed 27 days ago
https://github.com/moonshotai/kimi-cliAn AI agent for the terminal designed to assist with software development and terminal operations. Kimi CLI can read and edit code, execute shell commands, and search the web. It supports the Agent Client Protocol (ACP) for integration with IDEs like Zed and JetBrains, and provides a Zsh plugin for shell integration. The repository also includes the kimi-sdk for Python, the kosong library for agent orchestration and tool calling, and PyKAOS for OS interaction abstraction.
ACPKaos is a variant of LocalKaos designed to redirect specific operations to an Agent Client Protocol (ACP) client (such as Zed) while delegating all other operations to the local filesystem. This allows ACP clients to observe and participate in file edits and terminal command executions performed by the agent.
Key Redirections:
exec calls are redirected to ACP terminal operations.readtext calls are redirected to ACP fs/read_text_file.writetext calls are redirected to ACP fs/write_text_file.readTextFile and writeTextFile capabilities are supported; otherwise, it falls back to LocalKaos.Important Limitation:
Because stat, exists, and is_file checks currently use LocalKaos, the system may not recognize files that exist only in an ACP editor's unsaved buffers. These files will return false for existence checks until they are saved to the local disk.
Kimi Code CLI uses two distinct extension mechanisms:
SKILL.md file. The AI reads and follows these specifications (e.g., code styles, workflows, best practices).plugin.json. The AI can directly invoke these tools to perform actions like API calls, scripts, or database queries.Use Skills for defining how the AI should think or act, and Plugins for giving the AI tools to perform tasks.
Kimi Code CLI provides three primary ways to interact with the AI agent:
kimi): Chat with the AI in the terminal using natural language or execute shell commands directly.kimi web): A graphical interface in your local browser featuring session management, file references, and code highlighting.kimi acp): Run as a service to integrate with IDEs and other local agent clients via the [Agent Client Protocol].Kimi CLI includes several built-in tools for agents.
Default Tools:
Task, SetTodoList, Shell, ReadFile, Glob, Grep, WriteFile, StrReplaceFile, SearchWeb, FetchURL.
Optional Tools (must be enabled in Agent file):
Think, SendDMail, CreateSubagent.
Security & Approvals:
diff preview for user review.--yolo or --auto-approve flag to bypass manual approval prompts.Kimi CLI is designed as an AI-shell that provides an agent kernel through multiple communication modes. Beyond the standard terminal interface, it supports:
stdio to accept user prompts and push agent behavior events via a specific format. This mode powers the internal Web UI and VS Code extension.stdio) that allows Kimi CLI to integrate with any ACP client, including JetBrains, Zed, DeepChat, and Alma.Users can also extend Kimi CLI using MCP tools and skills or customize the agent by providing specific agent files containing system prompts and tool lists.
AGENTS.md file, run the /init command. This allows Kimi Code CLI to analyze your project and generate the file, which helps the AI better understand your project structure and conventions./init--continue, --session, or the /sessions slash command to resume previous work./clear (or its alias /reset) to wipe the current conversation context./compact to compress the current context to save tokens/space.Environment variables allow you to inject keys or change endpoints without modifying your configuration files. This is ideal for CI/CD or switching environments.
Overrides are determined by the provider type:
type = "kimi": Use KIMI_* environment variables.type = "openai_legacy" or type = "openai_responses": Use OPENAI_* environment variables.Refer to the environment variables documentation for a full list of supported keys.
To start the Kimi CLI, run the following command in your project directory:
kimiUse the /setup slash command to configure your platform and model settings.
Use the /help slash command to view available commands and usage information.
YOLO mode automatically approves all tool calls to remove approval friction. It does not prevent the agent from asking clarifying questions via AskUserQuestion if a decision is ambiguous. When active, a yellow YOLO badge appears in the status bar.
To enable YOLO mode:
--yolo flag./yolo command.default_yolo = true in your configuration file for permanent activation.# Enable at startup
kimi --yolo
# Or toggle during runtime
/yoloWhen operating in Plan Mode, follow this structured workflow:
Agent(subagent_type="explore") to investigate first.Glob, Grep, and ReadFile for quick lookups.ExitPlanMode to present the final plan to the user for sign-off.