What it is
An unofficial interactive breakdown of Anthropic’s Claude Code architecture and tool/command system, compiled from publicly surfaced source code and labeled with an analysis date of March 31, 2026.
Gabriel’s notes
Claude Code Unpacked is an unofficial, interactive explainer that maps “what happens when you type a message into Claude Code,” including the agent loop, a catalog of tools and slash commands, and a section of “hidden” or feature-flagged capabilities. The site explicitly states it is not affiliated with Anthropic, and that it is based on “publicly available source code,” with an analysis date of March 31, 2026.
Quick take: This is the kind of internet artifact you get when modern software meets modern chaos: a leaked/accidentally-published codebase, a fast community reverse-engineer, and then someone turns it into a clickable museum exhibit. It’s useful (and funny), but it’s also… legally and ethically spicy.
Here’s the cleaned-up version of my original note: News reports indicate Anthropic accidentally exposed internal Claude Code source code via a packaging/release mistake around March 31, 2026, and this site appears to be a snarky, community-built breakdown derived from that material. I genuinely like Anthropic and Claude Code, but this episode reads like a case study in “your build pipeline is also part of your security perimeter.”
I saved this under Dev & code because it’s a practical, concrete look at how an agentic coding CLI is wired—tools, loops, orchestration, and all the weird little UX decisions that never make it into a launch blog post.
Good fit if you want to:
- Understand the high-level “agent loop” behind a terminal-based coding agent (without squinting at a repo for 6 hours).
- Browse a structured inventory of tools/commands and see how they’re grouped.
- Learn how multi-agent orchestration is commonly implemented (worktrees, task fan-out, aggregation patterns).
- Sanity-check your own agent designs (permissions, tool boundaries, UI hooks, background execution patterns).
- Get a reality check on how much product behavior lives in prompts, flags, and glue code.
Pricing snapshot (auto-enriched):
Free to access as a website (no pricing page or paywall was visible at the time of review). Anything beyond that is Unknown / not confirmed.
Work-use / compliance snapshot (auto-enriched):
This is where you should slow down. Axios reported Anthropic said an internal Claude Code release included internal source code and that no sensitive customer data or credentials were exposed; it also noted takedown activity as the community reverse-engineered and re-hosted material. If your employer/client has a conservative legal posture, treat this site as informational only: don’t copy/paste code, don’t re-host artifacts, and don’t treat “leaked implementation details” as a safe dependency. When in doubt, loop in counsel—because “but it was on the internet” is not a compliance strategy.
Alternatives (auto-enriched):
- Anthropic’s official Claude Code docs: less gossip, more accurate/authorized setup and usage guidance—better for day-to-day work and safer to cite in client deliverables.
- Aider (open-source terminal pair-programming): if what you actually want is a strong CLI coding workflow (not a forensic tour of Claude Code internals), Aider is purpose-built for that and is openly licensed.
Before you adopt it:
- Decide your intent: learning and architecture inspiration is one thing; operationalizing leaked details is another.
- Document what you learned in abstraction (patterns), not in replication (implementation).
- If you’re sharing internally, add a one-liner: “Do not copy or redistribute any leaked source; use for conceptual learning only.”
Sources:
- https://ccunpacked.dev/
- https://www.axios.com/2026/03/31/anthropic-leaked-source-code-ai
- https://docs.anthropic.com/en/docs/claude-code/quickstart
- https://github.com/Aider-AI/aider