Kimi Code

• Recording a UI bug or reference interaction and asking the agent to reproduce or implement it from the video
• Running the CLI in a repository to refactor code, run the test suite, and fix failures
• Using the desktop app to work across several local projects, with a diff panel after each task
• Driving Kimi Code from Zed or a JetBrains IDE over ACP without a separate login
• Pointing Claude Code or OpenCode at the Kimi Code endpoint to use K3 through a Kimi membership

Kimi Code's case is breadth on one model family. The same agent runs in an open-source CLI, a desktop app, VS Code, and ACP editors, and the membership quota even works from Claude Code or OpenCode. Its most unusual feature is video input, and K3's 1M context is available on higher tiers. The tradeoff is that plan structure matters. Kimi Code starts at Plus, the full context and HighSpeed model need Pro, quota is shared with Kimi chat under rolling limits, and the tiers were just renamed. It is easiest to justify if you already want Kimi models; teams that switch vendors often will get more from a provider-neutral agent.

About Kimi Code

Kimi Code is the developer service included with Kimi membership, and it runs on the same quota as the Kimi chat app. There are four official clients. The CLI is MIT licensed and installs as a single binary with one script on macOS, Linux, or Windows (Windows also needs Git for Windows), or from npm. The desktop app runs on macOS 12+ (Apple Silicon and Intel) and Windows 10+ and adds workspaces, a diff panel after each task, and a built-in terminal and browser. The VS Code extension adds a chat panel with @ file references, diffs with rollback, and MCP servers. Zed, JetBrains, and other Agent Client Protocol editors can drive the CLI through `kimi acp`. The agent can dispatch built-in coder, explore, and plan subagents in isolated contexts, and it accepts video input, so you can drop in a screen recording and ask it to build what it shows. MCP servers can be added and authenticated conversationally with /mcp-config. Skills, MCP servers, and data sources install from a marketplace or any GitHub repo, with each install's trust level shown up front. Lifecycle hooks can gate risky tool calls or trigger your own automation. The models are k3 (up to 1M tokens of context), k3-256k, kimi-for-coding (K2.8 Preview), and kimi-for-coding-highspeed (K2.7 Code HighSpeed). The Kimi Code plan also exposes OpenAI- and Anthropic-compatible endpoints, and Moonshot documents using them from Claude Code, OpenCode, Codex, and Hermes Agent. You sign in with Kimi Code OAuth or a Moonshot AI Open Platform API key, and the CLI and desktop app can also be configured for other compatible providers. The main tradeoff is billing. Kimi Code is included only with Plus and higher memberships (the Go tier has no coding quota), K3's full 1M context and the HighSpeed model require Pro or higher, usage is limited by a rolling five-hour window plus a monthly quota shared with Kimi chat, and Kimi recently renamed its tiers. It suits developers already using or evaluating Kimi models. Teams that want one bill across many model vendors may prefer a provider-neutral agent.