Breaking

Open-source rebuild recreates Claude Code internals in 23 steps

June 28, 2026 at 21:09 EDT

AI Agents
Software Dev & Coding
Open Source

An open-source implementation that reconstructs the architecture of Claude Code from a minimal agent loop has been released. Shared on X by developer Tom Dörr, the GitHub repository claude-code-from-scratch recreates how Anthropic's terminal-based coding assistant works through 23 progressive Python scripts, from s01_perception_action_loop.py to s23_worktree_advanced.py.

Open-Source Recreation · Harness Engineering

Claude Code, Rebuilt From Scratch in Python

A clean-room reimplementation reverse-engineers Anthropic's autonomous CLI coding agent — growing from a minimal while-loop into a production-grade system of 23 components. It contains none of Anthropic's original code.

progressive session files, s01 → s23

392

lines in the shared core.py engine

186

GitHub stars since release

forks already created

"98% is not the model"

Claude Code's success is attributed less to the model than to the harness built around it. To scale: 1 block = the model's raw inference; the tall column = everything else.

~2%

Model call

~98%

The harness

The Core Loop

An AI agent is essentially a while loop with a model call — call the model, execute tools, repeat.

Call model

→

Execute tools

→

Repeat ↺

What the harness adds

Permissions

Declarative YAML-based governance

Context

Three-layer compression, KV / prompt caching

Multi-agent

Subagent isolation, FSM-protocol teams

Task isolation

Git worktree working-directory isolation

Execution

Parallel tools, MCP, Redis pub/sub mailboxes

Extension

.md skills, interrupts, streaming, background tasks

Praised for

Open, replicable production patterns
Self-correcting loops & sub-agent isolation
Clarity of declarative rule design

Watch for

Full harnesses add complexity
Long sessions hit rate limits & cost
Permissions need careful configuration

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.