getkimchi/kimchi

What it solves

Kimchi is a terminal-based coding agent that provides an AI-powered development assistant directly in the CLI. It streamlines the software development lifecycle by automating code implementation, codebase exploration, and research, while offering advanced project management for long-term tasks.

How it works

Built on the pi-mono SDK, Kimchi connects to a specialized LLM infrastructure. It uses a multi-model orchestration system where a primary orchestrator delegates tasks to specialized roles (Builder, Reviewer, Explorer, Researcher, and Planner) based on the complexity of the task and the model's capabilities. It integrates with Language Server Protocol (LSP) for type-aware intelligence and uses a state-machine-driven project mode called "Ferment" to persist structured plans across sessions.

Who it’s for

Developers who prefer working in the terminal and want an AI agent capable of handling complex, multi-step coding tasks with persistence, codebase awareness, and the ability to move sessions to cloud sandboxes.

Highlights

• Multi-Model Orchestration: Dynamically routes tasks to the best-fit model (e.g., using a heavy-tier model for complex algorithms and a light-tier for simple tasks).
• Ferment Project Mode: A progressive-refinement system that persists goals, phases, and steps in a JSON state file to resume work across sessions.
• LSP Integration: Built-in support for TypeScript/JavaScript and Go, providing the agent with real-time diagnostics, symbol definitions, and references.
• Remote Sessions: The /teleport command allows users to move an active session to a cloud sandbox for persistence or team hand-off.
• Context Management: Supports global and project-level AGENTS.md files for custom instructions and rules.
• Token Optimization: Integrates with RTK to compress bash command output, reducing LLM context usage by 60-90%.