Kimi Code CLI

Why it's ranked here

It stands out in the terminal-agent category on architecture rather than feature count alone. The core runs on a real dependency-injection framework with its own service lifecycle, not a pile of global functions, and the editor integration is a full implementation of a standard agent protocol, covering session forking, resuming, and permission bridging rather than a thin wrapper. Both point to a team building for more surface than a single command-line tool, which is why the same runtime works as a terminal app and an editor backend. The permissive license removes the usual proprietary-agent adoption friction.

What's good

The install path is genuinely light: a shell script drops a working binary with no separate runtime, PATH conflicts, or global package collisions to manage. The editor integration goes deeper than a status-bar plugin: sessions can be listed, resumed, closed, deleted, and forked mid-conversation, with approval prompts and plan review routed back through the editor. Lifecycle hooks let a team gate or audit any tool call before it runs, which matters once an agent can execute shell commands on its own. Outbound requests route through a single proxy setting, so it works behind a corporate firewall without code changes.

Tradeoffs

Windows setup carries an extra step: the agent shells out through Git Bash rather than a native shell, so first-time Windows users need a separate Git install before anything runs. The provider story defaults hard to the maker's own models; pointing it at another compatible provider means configuring that yourself rather than picking from a first-class list. The core package's internal surface is large and deeply layered, built on its own dependency-injection and service-lifecycle machinery, which is a lot of architecture for a newcomer to read before contributing a change instead of just calling a function.

How to use it well

It fits best as a daily driver for people who want to stay inside an editor: point a supported IDE at it through the editor protocol and get the same session, approvals, and plan review there instead of a separate terminal pane. Teams that want a record of what the agent touches should wire lifecycle hooks to their own audit or notification system before allowing unattended runs. On a large, unfamiliar codebase, dispatching the exploration and planning subagents first keeps the main conversation focused while the agent maps the territory before it starts editing. It is not the choice for someone who wants equal, first-class support across many model providers.

Observed

License

MIT licensed

Packaging

Distributed as a single native binary installed via a shell script, with an alternate npm install path

Interfaces

Interfaces: interactive terminal UI, headless prompt mode, an Agent Client Protocol server for editor integration, and MCP client support

Platform requirement

Requires Git Bash as the shell backend on Windows

Packaging

Organized as a monorepo of separate packages: CLI app, agent core runtime, ACP server, OAuth/auth, and a cross-platform shell abstraction layer

Lifecycle hooks

Ships lifecycle hooks that run local commands at defined points around tool calls

Input support

Supports video file input in addition to text prompts

Networking

Outbound requests are routed through a configurable HTTP/HTTPS proxy dispatcher