Kimi Code

The agent works directly in the developer's environment. Built in tools read, write and edit files, search with ripgrep and glob patterns, run shell commands in the foreground or in the background, search the web and fetch pages. MCP servers connect as local stdio commands or remote HTTP URLs, and plugins bring in more servers, skills and agents. The official Kimi Datasource plugin queries market, filing, macroeconomic, academic and Chinese legal data on request, read only, and allow and deny rules can be set per tool, including wildcards for MCP tools.

Sourcekimi.com/code/docs/en/kimi-code-cli/reference/tools.htmlread 2026-11-07

The main agent dispatches subagents with their own context windows, the built in coder, explore and plan agents or custom ones defined in Markdown, and several can run in parallel or in the background and be resumed later. Swarm mode launches many subagents at once.

The experimental Tower mode splits a larger task into missions, starts workers whose dependencies are met, gives each mission its own branch and worktree, sends finished branches to read only reviewers and merges approved ones in dependency order, with up to 256 workers and reviewers running together by default. /goal keeps a standing objective going across turns.

Sourcekimi.com/code/docs/en/kimi-code-cli/guides/tower.htmlread 2026-11-07

Context comes from the project as the agent works. /init analyzes the codebase and writes an AGENTS.md, and the agent reads files with line ranges, searches contents with ripgrep, finds files by pattern, reads images and video, and fetches web pages. Long sessions compact their history automatically or on /compact with a hint about what to keep. Kimi Datasource answers questions from outside market, filing, academic and legal sources, while no index or embedding store is built over the customer's own code or documents.

Sourcekimi.com/code/docs/en/kimi-code-cli/reference/slash-commands.htmlread 2026-11-07

Always Ask is the default mode. Reads run on their own, while every file write, edit, shell command and scheduled task waits in an approval panel where the user approves once, approves for the session or rejects. Ask When Needed lets routine calls through but still stops for sensitive files and dangerous commands, and Plan mode makes the agent draft a plan and wait before changing anything, with exit confirmed in every mode. Subagent dispatch needs approval unless a rule allows it. Never Ask runs everything unattended, and Tower workers always run in that mode.

Sourcekimi.com/code/docs/en/kimi-code-cli/guides/interaction.htmlread 2026-11-07

Kimi Code runs on individual Kimi accounts that sign in with an OAuth device code or a Kimi Platform API key. The local server requires a bearer token and blocks a source after 10 failed attempts in 60 seconds on network binds, and hook rules load only from user config and local plugins, never from a checked out repository. There are no organization roles, SSO, SAML, SCIM or audit log for an administrator, no SOC 2 or ISO 27001 report is named, and team and enterprise use is pointed to the separate Kimi Platform.

Sourcekimi.com/code/docs/enread 2026-11-07

Every session keeps an event stream for each agent and subagent, with a request trace used for recovery, replay and debugging. A session exports as a ZIP with its diagnostic logs or as a Markdown transcript, the local server returns paged message history and streams events such as tool calls and tool results, and /usage shows token and quota use. Hooks fire before and after every tool call, on each permission request and result, and on subagent start and stop, so a team can write its own log. There is no organization level export or central audit log.

Sourcekimi.com/code/docs/en/kimi-code-cli/guides/sessions.htmlread 2026-11-07

Each session is saved by working directory with its metadata and an event stream for every agent, so work can be resumed with --continue or by ID, forked into an independent copy, or compacted when the context fills. Recent Claude Code and Codex sessions can be imported, with the full history kept for the agent to consult. Subagent state is saved inside the session, and scheduled tasks reload with it. The CLI keeps no memory that carries facts from one session into the next.

Sourcekimi.com/code/docs/en/kimi-code-cli/guides/sessions.htmlread 2026-11-07

Kimi Code runs on the developer's own machine, as a desktop app for macOS, Windows and Linux, a CLI installed by script or npm, or a VS Code extension, and sessions can also bind to a remote environment. The CLI points at any provider the customer configures, including OpenAI compatible servers it hosts itself, so code and prompts can stay on its own infrastructure. Kimi models are served from separate China and overseas endpoints. Remote Control keeps every task on the user's machine and uses the web page only as a window.

Sourcekimi.com/code/docs/en/kimi-code-cli/configuration/providers.htmlread 2026-11-07

A plugin manager installs packaged capabilities from Official, Curated and Custom tabs or a marketplace file. The official catalog offers Kimi Datasource for financial and research data, a browser extension plugin and a computer use plugin, and Curated holds plugins from Kimi partners. Each plugin can bundle Agent Skills, custom agents, MCP servers, hooks and slash commands. The CLI also ships ready coder, explore and plan subagents and built in skills for configuration, themes and importing a Claude Code or Codex setup.

Sourcekimi.com/code/docs/en/kimi-code-cli/customization/plugins.htmlread 2026-11-07

Scheduled tasks start work without a person at the keyboard. Asked for a reminder or a recurring job, the agent writes a cron expression in local time and injects the prompt back into the same session each time it fires, such as summarizing CI failures every weekday at 9am. Tasks stay with their session and reload on resume, and recurring ones expire after 7 days. Everything else is started by a person, through the desktop app, the CLI, VS Code, JetBrains or Zed over ACP, the local server API, or Remote Control from a phone, and no webhook or CI event starts a run.

Sourcekimi.com/code/docs/en/kimi-code-cli/guides/use-cases.htmlread 2026-11-07

In the CLI, the customer adds providers of the Kimi, Anthropic, OpenAI, OpenAI Responses, Gemini and Vertex AI types, or any OpenAI compatible service such as DeepSeek and Qwen, and switches with /model, with a separate default for subagents. Desktop and VS Code pick among six Kimi model IDs, from K3.1 Pro and its HighSpeed version to K3, a 256K version of K3 and the K2.8 and K2.7 coding models, and higher tiers unlock the larger context and faster models. Choosing another maker is a CLI setting.

Sourcekimi.com/code/docs/en/kimi-code-cli/configuration/providers.htmlread 2026-11-07

A local server started with kimi web exposes sessions, prompts, approvals, configuration, models, providers and Remote Control over REST, with a WebSocket stream for turn, message and tool events, behind a bearer token. Live OpenAPI and AsyncAPI specs are served by the server itself, and Moonshot labels the API experimental. The CLI also runs as an ACP server for JetBrains, Zed and other editors, and a plugin manifest format lets teams package their own skills, agents, MCP servers and hooks.

Sourcekimi.com/code/docs/en/kimi-code-cli/reference/server-api.htmlread 2026-11-07

In Tower mode, each finished worker branch goes to an independent reviewer that inspects it read only, and the tower merges only branches that pass its checks. Plan mode puts an approach in front of the user before any file changes, /undo rolls back recent prompts, and the debug export packages a session's trace for diagnosis. There is no evaluation harness, set of test cases or scored comparison for checking a change to an agent, a skill or a prompt before it is used.

Sourcekimi.com/code/docs/en/kimi-code-cli/guides/tower.htmlread 2026-11-07

The Desktop app has a built in browser that the agent can operate itself, reading page text and elements, taking screenshots, clicking, typing, filling forms, choosing options and scrolling, once the user allows it, with a Take over control to stop it at any point and a separate setting for subagents. The official Kimi Browser Extension plugin lets the agent operate the user's everyday Chrome or Edge, and the official Kimi Computer Use plugin operates desktop apps on macOS and Windows, where the Windows version can briefly take over the mouse and keyboard.

Sourcekimi.com/code/docs/en/kimi-code-desktop/built-in-browser.htmlread 2026-11-07