Summary

Kimchi is a terminal-based AI coding assistant leveraging multi-model orchestration to provide intelligent development support. It empowers developers by connecting to LLM infrastructure directly within their workflow, offering features for task delegation, project management, and efficient code interaction.

How It Works

Kimchi employs a multi-model architecture where specialized LLMs are assigned roles like orchestrator, planner, builder, reviewer, and explorer. Tasks are delegated to these roles, and LLM requests are tagged with phases (explore, plan, build, review, research) for analytics. A core feature is "Ferment," a persistent project management system that breaks down work into goals, phases, and steps, managed by a deterministic state machine for robust lifecycle control across sessions.

Quick Start & Requirements

Installation is available via Homebrew (brew install getkimchi/tap/kimchi) or a curl script (curl -fsSL https://github.com/castai/kimchi/releases/latest/download/install.sh | bash). Post-installation, kimchi setup configures API keys interactively. The agent is launched with kimchi. For development, Node.js 22 (LTS), Bun, and pnpm are required. The rtk tool can be installed separately and integrated for significant LLM context reduction.

Highlighted Details

• Multi-model Orchestration: Dynamically assigns LLM roles for specialized task handling, enhancing efficiency and quality.
• Ferment Project Management: Enables cross-session, structured project planning with state persistence, recovery, and a validated state machine.
• Remote Teleport (Preview): Facilitates session multiplexing, allowing remote workers to be spawned and managed from a local terminal.
• RTK Integration: Compresses shell command output by 60-90% via rtk rewrite, drastically reducing LLM context window usage and costs.
• Customizable Hooks: Supports user-defined Bash hooks to intercept, rewrite, or block shell commands before execution.
• Agent Migration: Offers seamless migration of MCP servers and skills from Claude Code and OpenCode installations.

Maintenance & Community

The project utilizes GitHub Actions for automated releases upon version tagging. Specific community channels or maintainer details were not explicitly detailed in the provided README.

Licensing & Compatibility

Licensed under the Apache License 2.0, which permits commercial use and integration into closed-source projects.

Limitations & Caveats

The "Teleport" feature is currently in preview. Project-specific hooks are disabled by default and require explicit enabling. The agent's model support is dynamically fetched, potentially creating a dependency on external service availability. Development requires a specific Node.js/Bun/pnpm environment.