The Private Agentic AI Kit — build an AI assistant that provably never phones home: a local model, a local database, and a 3-tool agent loop in ~200 lines, with the network check that proves the zero
The companion kit to the video. We built a fully-local agentic assistant — Ollama running qwen3.5:9b (thinking OFF) as the brain, nomic-embed-text + one SQLite file (brain.db) as the memory, and three tools it calls on its own (search_documents, save_memory, list_memories) as the hands — then watched every network connection while it worked: 0 outside connections, on every snapshot, across both logged test batteries. INSIDE (8 pages): (1) THE 3-LAYER BLUEPRINT — Brain / Memory / Hands and exactly how the tool loop runs, with the diagram. (2) SETUP — four commands, plus the pick-by-RAM model table (8GB / 16-24GB / 32GB+), every model name verified live on ollama.com. (3) TRAP #1: THE THINKING-MODE TOKEN BURN — the default setting that spent 140 seconds, burned 4,035 tokens, and returned an EMPTY answer, and the one-line fix (think: false / /set nothink). (4) TRAP #2: MEMORY LOSES TO DOCUMENTS — the fresh session that trusted a stale document date over its own saved correction, and the fix: inject saved memories into the system prompt on every call, memories win by construction. (5) THE PRIVACY CHECK — the 7-line netcheck.sh, how to read it, and the zero that is the receipt. (6) FULL SOURCE of all three files — agent.mjs, ingest.mjs, netcheck.sh — zero packages to install. Independent kit from Hyperautomation Labs; not affiliated with Ollama or Alibaba (Qwen).
Free. No spam. Unsubscribe anytime.