An interactive study website that breaks down Claude Code's internal architecture, based on community analysis of the March 31, 2026 source leak.
🌐 Live: claude-code-study-five.vercel.app
On March 31, 2026, Claude Code v2.1.88 was published to npm with a 59.8MB source map file accidentally attached. This .map file exposed 1,902 TypeScript source files — the full internal architecture of Anthropic's agentic coding CLI.
This site organizes and explains the key findings for AI engineers and developers who want to understand how the leading agentic coding tool is built.
Disclaimer: This is educational material for learning purposes. It does not distribute proprietary source code. Data comes from the clawd-code community Python port which catalogs the architectural surface without reproducing the original TypeScript.
How Claude Code's runtime works under the hood:
- Bootstrap flow — The 7-stage startup pipeline (prefetch → trust gate → mode routing → turn loop)
- Single-threaded master loop (codename "nO") — Why Anthropic chose simplicity over complex multi-agent orchestration
- Tool system — 94 unique tools with permission layering, deny lists, and simple-mode fallback
- Command system — 141 slash commands from
/helpto hidden stubs like/bughunterand/teleport - Permission model — Trust-gated deferred init, sandbox toggle, destructive command warnings
🔮 Hidden Features
44 feature flags found in the source, with 20+ fully built but unshipped:
- Background Agents (Kairos) — 24/7 agents with GitHub webhooks and push notifications
- Multi-Agent Orchestration (Coordinator Mode) — One Claude spawning worker Claudes with restricted toolsets
- Ultraplan — 30-minute Opus session on remote server for deep task planning
- Voice Mode (Tengu) — Full push-to-talk with Deepgram Nova 3
- Browser Control — Playwright integration, not just
web_fetch - Cron Scheduling — CI/CD-style scheduled agent tasks
- The /buddy Pet System — An ASCII Tamagotchi with gacha rarity, shiny variants, and stats like DEBUGGING, CHAOS, and SNARK 🐤
Searchable, filterable catalogs of:
- All 207 command entries (141 unique commands)
- All 184 tool entries (94 unique tools)
Visual breakdown of all 28 subsystems with file counts — from utils (564 files) down to voice (1 file).
The human side of a $380B company's codebase:
- 460
eslint-disablecomments - 803KB
main.tsx(4,683 lines) - 50+
_DEPRECATEDfunctions still in production - Hex-encoded "duck" to dodge their own build scanner
- An engineer named Ollie openly questioning his own memoization in production
Practical takeaways you can apply to your own agent systems, including tool abstraction patterns, permission design, and multi-agent architecture strategies.
- Next.js 15 (App Router, static export)
- TypeScript 5
- Tailwind CSS 4
- Framer Motion (minimal animations)
- Deployed on Vercel
Architectural metadata is sourced from instructkr/clawd-code, a community Python port by Sigrid Jin that catalogs the structural surface of Claude Code's architecture. The JSON snapshots in src/data/ contain command names, tool names, subsystem structure, and source path hints — no proprietary source code.
git clone https://github.com/paulosalvatore/claude-code-study.git
cd claude-code-study
npm install
npm run devOpen http://localhost:3000.
This is a personal study project, but PRs are welcome for:
- Corrections to architectural descriptions
- Additional insights from community analysis
- UI/UX improvements
- New sections or deeper dives into specific subsystems
Built by Paulo Salvatore with the help of PaulinhoClaw 🤙
MIT — educational use. This project does not claim ownership of any Anthropic intellectual property.