Building an Agentic OS with Claude Code and Jev

The Breakthrough

The author implements a three-layer agentic operating system that uses Jev as a high-speed, low-cost classifier to route tasks between local models, Claude Code, and Codex, significantly reducing latency for simple requests.

What Actually Worked

• Tiered Model Routing: The system uses Jev to classify requests into three tiers: Tier 1 (simple commands, <100ms response) bypasses LLMs entirely; Tier 2 (non-complex tasks) uses smaller models like Haiku or Luna; Tier 3 (complex work) triggers Claude Code or Codex in the terminal.
• Skill Backbone: The author generates custom automation skills by feeding 30 days of Claude Code or Codex conversation logs into the model, prompting it to identify repetitive tasks suitable for automation.
• Structured Memory (RAG): The system utilizes a folder structure inspired by Andrej Karpathy, consisting of three primary directories: raw (source data), wiki (organized knowledge), and outputs (deliverables), which ensures the model can navigate the vault efficiently.
• Local Voice Integration: The system uses Whisper for local speech-to-text transcription and Coqui TTS for local text-to-speech, enabling a voice-controlled interface that functions independently of cloud-based voice APIs.

Context

The author addresses the inefficiency of using AI through standard terminal interfaces by building a custom dashboard (via web app or Obsidian plugin) that consolidates metrics, research, and task execution. The goal is to move beyond simple chat interactions toward a persistent, agentic environment that organizes personal or business workflows into repeatable, automated skills.

Content References

• Tool: Claude Code, Anthropic, mentioned
• Tool: Codex, mentioned
• Tool: Jev, TypeSafe, mentioned
• Tool: Obsidian, mentioned
• Tool: Whisper, OpenAI, mentioned
• Tool: Coqui TTS, mentioned
• Person: Andrej Karpathy, cited