A local LLM successfully replaced Claude as the brain of a personal AI agent after a hardware upgrade, according to a test replaying 27 real tasks. On a single RTX 3090 capped at a 16K context, a 30B model scored 22.8/100 against Claude’s 89.4 and leaked malformed tool-call syntax into a quarter of its answers. On three 3090s with a 256K context, a 122B model scored 80.0, eliminated malformed calls entirely (0 of 27), and cost $0.000969 per task versus Claude’s $0.763—roughly 787× cheaper. The agent, named Jarvis, is a LangGraph react-agent wired into roughly 90 tools including email, calendar, notes, files, Office, WhatsApp, Discord, and image generation. The test used 28 real historical task prompts from Jarvis’s Langfuse traces over a 90-day window, stratified across 7 categories with 4 each. Claude’s baseline was its actual historical production answer, frozen and identical in both rounds. The local model was re-run through a sandboxed harness with write-capable tools intercepted. The author cancelled Claude Max and kept Pro, noting the model size and context budget changed together with no control, so the result measures a system upgrade, not an isolated model effect.