Kimi Code Draft

Moonshot AI · reviewed 22 November 2026 · its card

Kimi Code lets a developer hand coding work to an agent that can "read and edit code, run shell commands, search files, fetch web pages" and choose its next step from the feedback: "an AI coding agent that runs in your terminal" (S1). Moonshot AI makes it, as MoonshotAI/kimi-code, under the MIT License; it presents it as "the intelligent coding service for developers among Kimi membership benefits, based on Kimi's latest flagship model" (translated from Chinese) (S1, S2). It runs as Kimi Code CLI in the terminal, a browser interface started by kimi web, Kimi Code Desktop on macOS, Windows and Linux, a VS Code extension, and in Zed and JetBrains through ACP (S1, S2, S14, S15). It replaces the Python Kimi CLI, Apache-2.0 and archived on 23 September 2026 (S19). The latest releases seen are Kimi Code CLI 2.7.2 and Kimi Code Desktop 1.4.2, both of 20 November 2026 (S3). It is an agent harness and a client.

An instance of
Agent harness, Client
Conforms to
Model Context Protocol (MCP), Agent Client Protocol (ACP)

How to read a product card.

Capabilities

A product page in four labelled parts: what it is and who makes it; the capabilities table, with columns term, value, detail and sources, and the values yes, partial, no and unknown shown by the matrix's glyphs; the state of play, dated, newest first; and the sources, each with the date it was read.
How to read a product page. Every product has the same parts in the same order.
CapabilityValueDetailSources
Model Yes Moonshot's own: k3.1-pro, k3, kimi-for-coding and others by membership tier; "Kimi K3 works best with Kimi Code CLI as its agent framework". Other providers: kimi, anthropic, openai ("OpenAI and compatible services, DeepSeek, Qwen, etc."), openai_responses, bedrock, google-genai, vertexai, and the models.dev catalogue; local models not named S2 S18 S12
Session Yes "persists every conversation as a "session""; kimi --continue, kimi --session <id>, /sessions (alias /resume) and forks S5
Subagent Yes "Dispatch built-in coder, explore, and plan subagents in isolated contexts"; custom Markdown agents; "Sub-agents support running in the background" S1 S6
Agent team Partial Tower, "an experimental multi-agent collaboration mode", off by default: the main agent "becomes the tower and coordinates these workers", each in its own worktree; "The tower routes messages", with no direct worker-to-worker messaging described S7
Skill Yes "SKILL.md consists of two parts: YAML frontmatter and a Markdown body", from .kimi-code/skills/, .agents/skills/ and user folders; the page is titled "Agent Skills" but does not name the specification S8
Instruction file Yes AGENTS.md or .kimi-code/AGENTS.md in the project, $KIMI_CODE_HOME/AGENTS.md and ~/.agents/AGENTS.md globally, "injected into the prompt as reference data" S6
Plugin Yes Plugins "can add Agent Skills, custom agents", hooks, commands and MCP servers; installed from "a local directory, zip URL, or GitHub repository URL" or the official marketplace S9
MCP Yes stdio, HTTP and legacy HTTP+SSE servers, OAuth by /mcp-config login, user and project mcp.json S10
Execution environment Yes container: remote environments run tools on an SSH host or "in a Docker-compatible container", with "no silent fallback to your machine"; no local sandbox documented, permission modes manual, yolo and auto S13 S11
Client Yes Terminal; web, on phones over LAN or Remote Control (kimi rc); Desktop; VS Code; Zed, JetBrains and Paseo through ACP S2 S14 S15
Agent host Unknown Desktop 1.4.0 lets "external agents such as Claude Code, Codex, and Cursor" connect to hand off sessions; nothing shows it running their sessions S3 S16
Channel Unknown No tool for outside events documented; scheduled tasks "re-inject a prompt into the current session" from within S17
Cross-session messaging Partial Its built-in "Kimi Agent Protocol (KAP)" local server lets other agents "interact directly with your local Kimi Code", for session hand-offs and "multi-round plan discussions", through the use-kimi-code skill; no built-in tool sends between its own sessions S16 S17
ACP Yes As agent: "Kimi Code CLI speaks the Agent Client Protocol"; editors launch kimi acp "as a child process" S1 S15
AHP Unknown Not found None
A2A Unknown The "Agent to Agent" page names only KAP, not Agent2Agent S16
Bot Unknown Not found. Scheduling: CronCreate schedules a prompt, but schedules "are bound to the session" and "not carried into a brand-new session" S17

State of play

How things stand, newest first, as the card records it.

  • Orca's version 1.4.269 lets Orca bots run on Kimi Code; this is Orca's claim, not Moonshot AI's. S20
  • Kimi Code CLI 2.7.1 and 2.7.2 and Kimi Code Desktop 1.4.1 and 1.4.2: web and Desktop fixes. S3 S4
  • Kimi Code CLI 2.7.0 adds Amazon Bedrock and resuming sessions open in another client; Desktop 1.4.0 connects external agents such as Claude Code, Codex and Cursor, adds SSH hosts and the Qwen Coding Plan provider. S3 S4
  • Kimi Code CLI 2.6.0 persists permission rules and adds artifact missions to Tower. S4
  • The Python Kimi CLI repository was archived; users move to Kimi Code CLI. S19
  • Kimi Code CLI 2.0.0 adds the /desktop command to install the desktop app. S4
  • Version 0.39.0 adds experimental Tower mode for multi-agent orchestration. S4