Interessant3 #153 | Claude Code, Offline AI, European Heatwaves
By Duarte Martins••485 words
Claude Code is All You Need
Conor Dwyer treats Claude Code as a Swiss Army knife for projects, server ops, admin, and even creative writing.
“Vibe coding” - letting AI spin up apps from prompts - works surprisingly well under the right constraints, as shown by a SmartSplit clone that runs cleanly in 900 lines of PHP but collapses into bloat when left to over-engineer in Node.js.
The “autonomous startup” experiment illustrates both potential and absurdity: Claude autonomously built and deployed a SaaS-style web app, but with shaky business logic and eventual policy blockages.
For real-world use, Claude excelled at tasks like migrating a legacy PHP/MySQL app, auto-documenting codebases, handling obscure dependency errors, renaming and merging bank statements, and even helping draft this very article.
Yet, the human role remains central: steering, clarifying, forgiving mistakes. Claude is less “independent coder” and more “hyper-diligent junior dev” that thrives on abundant context.
Manish – I Want Everything Local: Building My Offline AI Workspace
Begins from a simple but radical demand: no cloud, no remote execution. The aim - LLMs, code execution, and browsing, all on-device.
The stack blends Ollama (local LLMs), Apple’s new Container tool for VM-level isolation, Playwright for headless browsing, and a lightweight orchestration layer called coderunner.
Privacy is the philosophy: code runs inside isolated containers, results are mapped to shared volumes, and the host system is never touched - eliminating the risk of sensitive data leakage to third-party APIs.
The UI story reveals the messy reality of “offline-first” AI: failed attempts at a Mac-native app, a fallback to Electron/Next.js, and workarounds for missing features in open-source tools.
Tool-calling support and interoperability remain patchy - some models are advertised as having tool support but don’t actually implement it yet. Workarounds rely on exposing everything via the Model Context Protocol (MCP).
Tested use-cases include local video editing, chart generation from CSVs, browser-based research, image manipulation, and GitHub tool installs - all AI-orchestrated but fully offline.
Limitations: Apple-only for now, fragile build processes, and bot-detection throttling the headless browser. Yet the project signals a deeper shift: reclaiming compute and privacy from the cloud giants.