That said, Claude Code’s source isn’t a complete mystery, and while this exposure gives us a look at a fresh iteration of Claude Code straight from the leaky bucket, it’s not blowing the lid off of something that was a secret until now.
Claude Code has been reverse engineered, and various projects have resulted in an entire website dedicated to exposing the hidden portions of Claude Code that haven’t been released to, or shared with, the public.
In other words, what we have is a useful comparison point and update source for the CCLeaks operators, and maybe a few new secrets will come to light as people dig through the exposed code.
Far more interesting is the fact that someone at Anthropic made a mistake as bad as leaving a map file in a publish configuration. Publishing map files is generally frowned upon, as they’re meant for debugging obfuscated or bundled code and aren’t necessary for production. Not only that, but as we’ve seen in this example, they can easily be used to expose source code, as they’re a reference document for that original.
Antropic, more open than Open-AI.
I’m not sure I get this joke, seeing it everywhere. Is it because Anthropic is the least open of the major labs?
Anthropic has no open weight models, frequently sends their lawyers, DMCAs anything these leaks touch (see the last time this happened), and are heavily restrictive on usage of their platform.
Meanwhile, OpenAI has released a handful of open models, contributes to public research, heavily sponsors open source, and allows you to do whatever you want with their Codex infra (also open source).
Using Codex app-server, you can let your own users sign in with ChatGPT and use their subs for your app calls to Codex on their behalf. Not allowed with Anthropic. Not even Anthropic’s SDK is open source, except in the case of Python.
The models are commodities. It’s the the harness, tool wiring, agent workflows, mem/token optimizations, hidden feature flags, etc that are the claude secret sauce. All now open.