Claude Code might be the most underrated programming agent right now.

Not because of the model itself — but because of the Agent Harness. The model is raw material; the harness is the moat.

Someone extracted the full source code via the npm source map and created a clean fork: free-code.

Key changes:

- Stripped all telemetry (OpenTelemetry, Sentry, and callbacks all removed)

- Lifted system prompt restrictions (the hidden security layer was removed)

- Unlocked 45+ experimental features (voice mode, multi-agent planning, deep thinking, IDE bridging……)

Compile it from source yourself, connect your own API key, and run it locally.


This project is actually quite impressive. But scrolling through the comments, I found someone who put it more clearly:

> Forking the code itself isn’t enough. The real moat isn’t in that set of code, but in the .claude/ folder — Skills, Hooks, CLAUDE.md, and memory files. These are the assets that accumulate with your project and grow thicker the more you use them.

> The new fork gives you the engine, but not the context you’ve accumulated. A newbie who installs it will have to start from scratch every time. A veteran using the same model can be ten times faster at getting work done.

To put it simply: Anyone can call the model, but how you feed it context, how you configure skills, and how you make memories persist across sessions — the configuration of this wrapper is the line that separates efficiency from sluggishness.

Over the past two years, everyone has been focused on model parameters and benchmark rankings, but the design of the orchestration layer might be the variable that determines whether AI can actually do the work for you.