Open-source Claude Code rewrite by Sigrid Jin

Claude Code's source code leaked yesterday. And what's inside is more interesting than anyone expected. The leak exposed internal architecture, codenames, and a feature called KAIROS. KAIROS is an always-on background agent that runs while you're idle. It consolidates memory, merges observations, and converts insights into facts automatically. This isn't a chatbot. It's an AI that gets smarter about your codebase while you sleep. Here's what the leak revealed: ☑ KAIROS agent runs in the background — learns your codebase without you asking ☑ Coordinator Mode: one agent spawns and manages multiple worker agents in parallel ☑ Internal codenames: Capybara and Fennec are unreleased Claude variants in testing ☑ 259 pull requests, 497 commits in 30 days — one developer, one Claude Code instance ☑ 46% of developers call it their most loved tool — Cursor sits at 19%, Copilot at 9% ☑ The architecture shows Claude Code was built to replace IDEs, not work inside them Most developers are still using Claude Code like a smarter autocomplete. The leak shows it was built for something much bigger. I put together a free Claude Code starter guide for developers. Comment 'CLAUDE' and I'll DM it. Are you using Claude Code for full autonomous tasks yet, or still just code completion? #ClaudeCode #AIEngineering #DeveloperTools #BuildingWithAI #Agentic

The 48-Hour Clone: Why the Anthropic Leak Changes Everything for Codebase Owners If you own, manage, or protect a proprietary codebase, the events of March 31st, 2026, should fundamentally change how you view your intellectual property and competitive moat. When Anthropic inadvertently exposed the source map for their Claude Code tool, the open-source community completely rebuilt it. Using AI orchestration, an independent developer executed a "clean-room" reverse engineering of a 500,000-line enterprise architecture in under 48 hours. They hit 50,000 GitHub stars in two hours, entirely without copying a single line of copyrighted code. This matters for your engineering strategy today: • Your moat is no longer your syntax; it is your architecture. In the past, clean-room reverse engineering was a grueling, highly expensive legal and technical process. Today, if an AI can analyze your structural patterns, API calls, and functional specs, it can recompile your entire system from scratch. They mapped the logic in Python, then hardened it into Rust—all over a weekend. If competitors can see your blueprint, they can now build your house. • Developers are now orchestrators. The claw-code project wasn't built by an army of human coders typing out loops and variables. It was driven by multi-agent AI tools like oh-my-codex (OmX), running parallel execution loops and automated parity checks. You are competing against developers moving at light speed. • Security mistakes are instantly fatal to exclusivity. A simple deployment error, accidentally bundling a debug source map into a public NPM package is not a mild embarrassment, it is an instant commoditization event. You must secure your CI/CD pipelines with zero-tolerance policies for exposed debug artifacts. • Complexity is no longer a defense. The open-source rebuild achieved 40/40 tool specification parity almost immediately. You can no longer rely on the sheer size, legacy weight, or complexity of your codebase to deter competitors. AI does not get fatigued by reading half a million lines of context. The paradigm of software engineering has officially shifted. We have entered an era where abstract ideas and exposed architectures can be compiled into hardened, production-ready enterprise systems in days. If you found this of interest please like, comment, reshare, follow #SoftwareEngineering #ArtificialIntelligence #TechLeadership or #OpenSource

🧠 Building an Agentic RAG System — Phase 6 🚀 6 phases. One push to production. Here's what it took to ship it. Phase 6 is the one that makes everything before it real. Here's what shipped: 🐳 Multi-stage Docker build — python:3.12-slim base, Torch and PyWin32 excluded entirely by leaning on Gemini and Jina HTTP APIs instead. Final image: <500MB. 🔁 GitHub Actions CI/CD — lint → test → deploy on every push to main. Ruff formatting gates, Pytest suite with env secrets, and dual curl webhooks refreshing both the API and UI on Render. PRs blocked from triggering deploys. 🖥️ Streamlit UI — conversational frontend talking to the FastAPI backend on port 8000. What was a set of API endpoints is now something you can actually use. ⚙️ Entrypoint automation — alembic upgrade head runs before every Uvicorn boot. Schema migrations are idempotent and automatic. 🐛 2 bugs worth knowing: → Qdrant's Alpine image has no curl, bash, or wget — Docker's healthcheck stalled forever waiting for a signal that could never come. FastAPI never started. Fixed by removing the healthcheck for qdrant image entirely. → All 64 Pytest tests failed instantly in GitHub Actions with ModuleNotFoundError. The runner had no way to resolve internal imports. One line fixed it: env: PYTHONPATH: . That's the full build. 6 phases, shipped. What this project covers end to end: PDF ingestion → hybrid search → cross-encoder reranking → LangGraph agent → multimodal Gemini → Langfuse observability → prompt A/B testing → Docker → CI/CD → live on cloud. If you've been following along — thank you. Every comment, question, and pushback made the build better. What should I build next? #BuildingInPublic #RAG #Docker #CICD #AIEngineering #LangGraph #GenerativeAI #FastAPI

𝗖𝗹𝗮𝘂𝗱𝗲 𝗖𝗼𝗱𝗲'𝘀 𝘀𝗼𝘂𝗿𝗰𝗲 𝗹𝗲𝗮𝗸𝗲𝗱 𝗼𝗻𝘁𝗼 𝗡𝗣𝗠 𝗮𝘁 𝟰 𝗔𝗠. 121,000 GitHub stars in 48 hours. Here's what actually happened and what developers did next: Someone pushed Anthropic's internal #Claude Code source to a public npm package. Not a clone. Not a wrapper. The actual compiled harness. The dev community lost its mind. And here's what most people got wrong: → They called it a "#hack" → They assumed it was stolen #maliciously → They downloaded it without understanding what they had → They missed the real story entirely → Then the repo got pulled and they blamed Anthropic That's not a leak problem. That's a reading comprehension problem. What actually happened was more interesting: One engineer woke up at 4 AM in Seoul. Read the exposed source. Understood the agent harness architecture. Rewrote the core in Python from scratch clean room before sunrise. Then ported it to Rust. No proprietary code. No legal liability. Just architecture intelligence. Here's what the claw-code repo actually revealed: → How Claude Code wires tools to an agent runtime → The MCP orchestration layer under the hood → How session state and prompt compaction actually work → Why the slash command system is more powerful than it looks → The plugin hook pipeline most developers never knew existed → Why Claude generates "cleaner, more shareable code" it's architectural, not magical What this is: a masterclass in reverse-engineering through observation. What this is NOT: a free pass to ship Anthropic's IP as your own product. The repo hit 121K stars because developers recognised something rare: Someone who understood the system well enough to rebuild it. That's the actual skill. Not downloading a leak. Understanding why it was built the way it was. Want the full breakdown of what the Claude Code architecture actually teaches us about building AI agent systems? 1. Follow me 2. Comment "𝐂𝐋𝐀𝐖" below I'll send you the full teardown. ♻️ Repost if you saw the repo trending and didn't actually understand what you were looking at. There are more of us than we admit.

brooks-lint v0.6.1 — now tri-platform! AI code reviews grounded in 6 classic engineering books now work on Claude Code Gemini CLI Codex CLI. 4 modes: PR review, architecture audit, tech debt, test quality. Every finding follows: Symptom → Source → Consequence → Remedy. New in v0.6.1: Codex CLI support with one-command install. GitHub: https://lnkd.in/gXus8itJ #CodeReview #AI #SoftwareEngineering #CodexCLI #ClaudeCode

Stop writing PR descriptions manually. The Eulogik team built AutoPR to automate it entirely using 100% free AI models. Here's the problem every developer knows: ❌ Writing PR descriptions is tedious and time-consuming ❌ Manual code reviews are slow and inconsistent   ❌ Context-switching from coding to docs kills flow So we built AutoPR – an open-source CLI + GitHub Action that: ✅ Auto-generates professional PR descriptions from git diffs ✅ AI-powered code reviews with security/performance checks ✅ Runs 100% FREE via OpenRouter (Gemma, Llama, Mistral) ✅ Works as CLI: `autopr generate --no-dry-run` ✅ Or fully automated GitHub Action on every PR Built with TypeScript strict mode, 22 passing tests, zero bloat. MIT license – free to use and modify. 🔗 GitHub: https://lnkd.in/dNt4UiYH 📦 npm: npm install -g @eulogik/autopr Would you use this? Drop a ⭐ if you're tired of writing PR descriptions! #OpenSource #GitHub #AI #CodeReview #DevTools #Automation #TypeScript #DeveloperProductivity #LLM #GitHubActions #Coding

Anthropic accidentally shipped the entire Claude Code source code to npm this morning. 512,000 lines of TypeScript. 1,900 files. Just sitting there. The way it happened is almost poetic. They built something called "Undercover Mode," a whole subsystem specifically designed to stop Claude from leaking internal codenames in git commits. Then shipped the entire source in a .map file. Likely packaged by Claude itself. Anyway. Here is what was inside. One of the pet species names collided with an internal model codename, so rather than rename it, someone hex encoded all 18 species names to dodge their own build scanner. There is a line in the codebase that reads: export const duck = String.fromCharCode(0x64,0x75,0x63,0x6b) That is the word duck. They hex encoded duck. main.tsx is 803,924 bytes. One file. 460 eslint-disable comments. A function called writeFileSyncAndFlush_DEPRECATED() still actively called in production. The error handler has a comment that reads "TODO: figure out why." An engineer named Ollie left a note openly admitting the code he shipped might be pointless. Look, I work in production AI systems. This looks exactly like every production codebase I have ever seen. The difference is nobody usually gets to look inside. A $380 billion company with the same 3am energy as the rest of us. Somehow that is both reassuring and terrifying. What is the worst comment you have ever left in production code? #AIEngineering #ClaudeCode #SoftwareEngineering

Anthropic built an AI to help you write better code. Then accidentally leaked 512,000 lines of their own code via a .map file in an npm package. 🧠💥 Security researcher Chaofan Shou spotted it. The internet did the rest — GitHub mirrors, 41,500+ forks, and a Hacker News thread that's still on fire. What was inside? → Full agent architecture & 40+ tools → A Tamagotchi companion called "Buddy" with CHAOS and SNARK personality stats (yes, really) → "Undercover Mode" — a system to hide Anthropic's internal codenames from git commits → KAIROS: an always-on background AI daemon that literally "dreams" while you're idle The root cause? Bun's bundler ships source maps by default. A missing line in .npmignore. That's it. As one dev put it: "A single misconfigured .npmignore can expose everything." Lesson for every engineering team: your build pipeline is a security surface. Audit it like one. Also — check your .npmignore. Right now. I'll wait. 👀 #WebDev #JavaScript #npm #AI #ClaudeCode #DevSecOps #BuildPipeline #Anthropic

I built my first AI agent from scratch today — and it actually works. Not with LangChain. Not with LangGraph. Raw Anthropic SDK in TypeScript. The agent analyses any GitHub repository end-to-end: → Fetches repo metadata via the GitHub API → Explores the file tree → Reads key source files (README, package.json, entry points) → Calls a structured submit_report tool when it's done The agentic loop is simple but powerful: plan → act → observe → repeat. Every tool call is a deliberate step — no magic, no abstraction hiding what's happening under the hood. Starting with the raw SDK was the right call. When you build the loop yourself, you actually understand what frameworks like LangGraph are doing for you — and when you'd want them. What I learned: → Structured output via a "finish" tool pattern is cleaner than parsing free text → Tool schema design matters — bad schemas = confused agent → Stateless, deterministic tool execution makes debugging simple Building toward a Next.js frontend demo next, then Lambda deployment. If you're learning AI engineering — build something real first. Skip the tutorials. link- #AIEngineering #TypeScript #LLM #AgentAI #BuildInPublic