Mini-Kode

About

Mini-Kode addresses the learning gap between toy agents (a few hundred lines) and production harnesses too large to read: at roughly fourteen thousand lines of TypeScript it is a working agent whose internals can actually be studied. The loop streams from any OpenAI-compatible endpoint (DeepSeek and GLM verified in the README), dispatches a unified tool system for files, search, and command execution behind two-layer permission approval, and reads AGENTS.md at startup for persistent project conventions — mirroring the conventions-file pattern of production agents. Architecture is deliberately legible: separate modules for tools, permissions, LLM, sessions, and UI, written for readers rather than throughput, with a DeepWiki walkthrough accompanying the code. A roadmap of session persistence, subagents, and context caching marks the gaps between it and production harnesses. Developers use it to learn how coding agents are assembled and as a base for their own experiments rather than for daily production work.