Third time in a year, so it's fair to ask whether Anthropic's leaks are a guerrilla marketing tactic or the result of vibed processes.👇
The thing I find puzzling about Anthropic leaking Claude Code's source on NPM [i] is that everyone is reacting with surprise.
It's not even about it being the second time in a week [ii] (and the third in a year [iii]), but rather about Anthropic being an outspoken advocate and practitioner of vibe coding, to the point that Claude Code is left to developing Cowork autonomously according to them, and by now we've all seen how that as an approach can lead to disaster in production environments when humans lower their level of alert and awareness, even at AWS. What's new? 🧐
Also, with an abundance of open-source coding agents in the wild, including those of Anthropic's biggest competitors, there's not much IP at risk here; we know by now what the anatomy of an agent looks like(a stack of custom prompts for the most), and all the real IP is in the models (and their training data) anyways; in fact, Anthropic just marked the "leaked" release on NPM as deprecated; it's likely still there.
What I am reacting to is rather the numbers really: 1,900 files and over 500K lines of code.
For a terminal app which is not much more than a wrapper for API calls and, as we now know from the leaked code analysis, a very thin orchestrator of subprocesses running other CLI tools, that's excessive.
A lot of the code has to do with Anthropic's hermetic choice to use React for rendering a terminal UI at 60fps [iv]. Well, that and... a lot of bad practices. 😅
Here's a pearl highlighted in a thread breaking down the leaked code [v]; src/cli/print.ts contains a core function for the agent loop, and it's quite representative of what a vibe-coded solution ends up looking like over several iterations when no human is paying attention anymore:
- 3,167 lines long (the file itself is 5,594 lines)
- 12 levels of nesting at its deepest
- ~486 branch points of cyclomatic complexity
- 12 parameters + an options object with 16 sub-properties
- Defines 21 inner functions and closures
- Handles: agent run loop, SIGINT, rate-limits, AWS auth, MCP lifecycle, plugin install/refresh, worktree bridging, team-lead polling (while(true) inside), control message dispatch (dozens of types), model switching, turn interruptionrecovery, and more. 😰
Going by that, at this point, Claude Code's codebase must be vibe-coded out of pure necessity, as clearly no human can be reasonably expected to navigate and maintain something like that.
Now I really understand why Claude Code is such a flaky, broken, and sluggish experience; I don't even want to know the stats for Claude Desktop and Cowork then. 🫣
Maybe this is why Anthropic didn't open source their client like everyone else running in this race has done already. 90% of the companies that have adopted agentic coding as part of their strategy haven’t figured out how to manage effectively the human+LLM SDLC at scale[vi], and Anthropic is just another player trying to move faster than its security and quality processes could possibly handle; at this stage there's no proof of good replacements for those processes yet that can enable LLM-boosted speed with reduced risk, which is what the "coding was never the bottleneck" choir has been pointing at for quite a while.
Being a business selling a new way of working as a cure to a productivity plateau in Tech, I can understand why they wouldn't want prospective customers to see what the actual side effects look like before those folks start self-administering the cure. Enterprise customers might not appreciate the seller suffering from its own adoption.
My bet is that keeping Claude Code closed source was never about protecting IP or building a moat; it was about hiding the chaos of unchecked vibe coding in an organisation that has skipped investing in the fundamentals: easily replicable environments, robust and fast CI/CD with a solid, thoroughly vetted, well-maintained testing suite, and clear security guidelines are the real "harness" that can unblock a lot of the delivery flow across an engineering organisation. Automate judiciously and invest in innovating there first, and both human and LLMs can fly at warp speed to production safely, to the benefit of customer satisfaction rather than in service to misleading code-related productivity metrics.
At least Anthropic is not going out in public positioning themselves as a centre of vibe coding excellence selling consultancy services to help businesses with the transformation, yet. 🎤⤵️👟➡️