Moonshot AI has published Kimi Code CLI on GitHub, a terminal-based coding agent built around its own Kimi models but configurable to run against other compatible providers. The tool installs as a single binary, distributed through a curl script, Homebrew, or a PowerShell command on Windows, with no separate Node.js setup for end users. It lands shortly after Moonshot’s Kimi K3 model launch, and the sequencing signals an attempt to own both the model layer and the agent layer that developers use to reach it.
The CLI opens into an interactive terminal interface Moonshot says starts in milliseconds. A developer logs in with either Kimi Code OAuth or a Moonshot AI Open Platform API key, then hands the agent a task, such as mapping out a project’s directory structure. From there the agent can modify files, execute shell commands, search a codebase, and pull in web pages, choosing its next action from what each step returns.
Three purpose-built subagents, labeled coder, explore, and plan, each run in an isolated context so a research detour does not clutter the main conversation thread. Lifecycle hooks let a team fire local scripts at defined points in a session: gating a risky file edit, logging a decision, or triggering a desktop alert. A marketplace-style plugin system adds skills, MCP servers, and data sources pulled from GitHub repositories, with each installation’s trust level shown before it runs.
Model Context Protocol, the standard Anthropic built for connecting agents to external tools and data, can be added and authenticated through a conversational /mcp-config command rather than hand-edited JSON files. The CLI also speaks the Agent Client Protocol, letting it act as the backing agent inside Zed or JetBrains through a kimi acp subcommand once a developer has logged in from the terminal.
The feature set (subagents, hooks, skills, MCP configuration, IDE protocol support) closely tracks the extension model Anthropic built for Claude Code. That overlap looks deliberate. A developer who already writes hooks or packages skills for Claude Code can port most of that work to Kimi Code with limited translation, which cuts the cost of trying Moonshot’s stack instead of committing to it. The repository does not publish independent benchmarks on task completion rate or session cost, so how the agent performs against Claude Code or OpenAI’s Codex on real repositories stays unverified outside Moonshot’s own description.
Teams already standardized on Claude Code’s skill and hook format should treat Kimi Code as a low-friction agent to benchmark this quarter, particularly for workflows priced against Kimi token rates rather than Claude’s. The open question is whether Moonshot backs the CLI with the same subagent reliability and hook ecosystem depth that made Claude Code sticky in the first place.
Details are drawn from Moonshot AI’s MoonshotAI/kimi-code repository and README published on GitHub.
