I want everything local – Building my offline AI workspace | AI Paper Digest

TL;DR Highlight

Building a fully offline AI workspace with LLM chat, code execution, and web browsing using Ollama, Apple Container, and Playwright — no cloud dependency. A meaningful attempt at replicating ChatGPT-level workflows while keeping personal data local.

Who Should Read

Developers who want to process sensitive personal data (photos, documents) with AI but don't want to upload to the cloud. Apple Silicon Mac users looking to set up their own local LLM environment.

Core Mechanics

Goal: replicate ChatGPT/Claude's 3 capabilities (LLM chat, code execution, web access) entirely locally. Uses Ollama for LLM, Apple Container (independent VM per container) for code execution, Playwright for browser automation.
Initially tried building a native Mac app with a0.dev but it was too iOS-focused, so pivoted to a Next.js web app — which worked much better for this use case.
Complete data privacy — nothing leaves the local machine. Personal photos, documents, and conversations stay private.
Hardware is the biggest bottleneck — 80B+ parameter models need large amounts of fast RAM. Even Mac Mini with sufficient RAM costs $2K+.

Evidence

Hardware was the most-discussed bottleneck. Running 80B+ models requires massive high-speed RAM — Mac Mini with enough RAM exceeds $2K. Framework desktop (128GB unified RAM) at $2-2.5K was mentioned as an option.
Berkeley research team's LEMO system was referenced as another open-source local AI workspace approach.
Apple Container provides VM-level isolation per code execution, addressing security concerns about running LLM-generated code.

How to Apply

For AI processing of personal photos/videos/documents without cloud exposure: Ollama + Apple Container keeps all data local. Just add the coderunner URL to mcpServers config for code execution from Claude Desktop too.
For local LLM tool-calling needs: Llama 3.3 70B via Ollama provides decent tool-calling support. Pair with MCP servers for file system, database, and web access.
Budget carefully — local AI that matches cloud quality requires significant hardware investment ($2K+ for a usable setup).

Code Example

snippet

// Configure coderunner MCP connection in Claude Desktop or Gemini CLI
{
  "mcpServers": {
    "coderunner": {
      "httpUrl": "http://coderunner.local:8222/mcp"
    }
  }
}

Terminology

OllamaA tool for easily downloading and running LLMs locally. Like Docker for AI models — manages model downloads and serving.

MCPModel Context Protocol. A standard interface for AI models to call external tools (code execution, DB queries, etc.). Like USB ports — various tools connect through the same standard.

Apple ContainerApple's containerization technology providing VM-level isolation. Each container runs in its own lightweight virtual machine for security.

I want everything local – Building my offline AI workspace