Half a million lines of TypeScript, compressed to 1,400 of Python

What it does

CoreCoder is a ~1,400-line Python implementation of the core patterns that make Claude Code work: search-and-replace editing, parallel tool execution, three-layer context compression, sub-agents, dangerous-command blocking, session persistence, and dynamic system prompts. It runs against any OpenAI-compatible API and ships as both a CLI and a library. The author reverse-engineered the leaked 512,000-line TypeScript source and kept only the load-bearing walls.

The interesting bit

The project treats minimalism as a feature, not a constraint. It intentionally omits Skills, MCP, and plugins so the codebase stays a teaching artifact rather than creeping toward yet another bloated agent framework. If you want the full tool, the author’s separate AnyCoder project exists; this is the annotated schematic.

Key highlights

• Seven Claude Code patterns implemented in runnable Python, none longer than 200 lines (context.py is the largest at 196).
• Model-agnostic by default via OpenAI-compatible APIs; optional LiteLLM extra unlocks 100+ providers including Bedrock and Vertex.
• Extensible tool system: custom tools inherit from a base class and run in about 20 lines.
• Includes session save/resume, mid-conversation model switching, and a /compact command for context compression.
• Seven accompanying articles dissect Claude Code’s agent loop, streaming executor, and 44 hidden feature flags.

Caveats

• Deliberately incomplete: no Skills, Subagents, or MCP support, so it’s a learning scaffold rather than a daily driver.
• If you want a practical tool, the author points you to AnyCoder instead.

Verdict

Grab this if you want to understand how Claude Code actually works under the hood; skip it if you just need an AI to edit your files and don’t care about the machinery.