Moonshot AI's open-source terminal AI coding agent kimi-cli is quietly getting a major overhaul — a full repository migration and architecture rewrite — and has been officially renamed kimi-code. To fix the original Python version's laggy response and execution inefficiencies, the research team has completely switched to the technical approach of Anthropic's terminal tool Claude Code. The result is a complete architecture rebuild using TypeScript and the Bun runtime, delivering millisecond-level cold starts and a smooth terminal user interface (TUI).
This architecture shift means Kimi has thrown out its original Python terminal stack entirely, adopting Claude Code's proven design. The tool now uses Commander.js for command parsing and React Ink — replacing Rich and prompt-toolkit — for a responsive new TUI. The rewrite involved 166 TypeScript source files and over 38,000 lines of new code. On the SWE-bench Verified benchmark, the TypeScript rewrite powered by the kimi-k2.5 model solved 317 out of 500 development tasks, a 63.4% solve rate. Performance matches the original Python version, while stability and network resilience have improved significantly.
Beyond the architecture, Moonshot AI has polished kimi-code's human-AI collaboration experience. The new version not only supports multimodal analysis by dragging screen recordings and other video assets directly into the terminal, but also deeply replicates several of Claude Code's signature features. These include a "Planning Mode" with cursor-based interactive editing, Emacs keybindings, a safety design that exits on double Ctrl+C, and support for custom lifecycle hooks to integrate with automation workflows. On multi-model compatibility, kimi-code opens up the ability to plug in third-party large model APIs, making it a unified terminal programming gateway across models, not just limited to Kimi's own.