SmartBrain

Blog

Engineering notes from building a self-hosted, local-first AI assistant.

Turning Claude Code into a pure LLM endpoint

September 7, 2026

We wanted a Claude subscription to work as a plain model backend — no API key. Driving a full coding agent as a bare model took real containment work: the obvious tool-blocking flag blocks nothing, your global CLAUDE.md rides into every session by default, and argv is public. What leaked, how we proved it, and the tested recipe that seals it.