I Built a Subscription Scanner in 18 Hours. It Runs on My Hardware.
Your email knows what you're paying for. We just ask it. Locally. No cloud. Sovereign.
<p>Nate Jones said it in his Mythos video this week: <em>"Just tell your AI to look at your recurring subscriptions and find $200 in savings."</em></p>
<p>So I built the tool that does exactly that.</p>
<h2>What It Does</h2>
<p>The Subscription Trimmer scans your Gmail for recurring charges — receipts, invoices, payment confirmations, renewal notices. A local LLM (Qwen2.5-72B running on my RTX PRO 6000 Blackwell) classifies each email: service name, amount, frequency, category. Then it identifies "ghost subscriptions" — services that charged you but you haven't engaged with in 30+ days.</p>
<p>The result: your total monthly burn, laid out in clean cards, with ghosts highlighted in amber.</p>
<p><strong>Live demo with real data from my own email: <a href="https://ganuda.us/trimmer">ganuda.us/trimmer</a></strong></p>
<h2>How It's Built</h2>
<p>Gmail API (OAuth, read-only) → Scanner pulls subscription-related emails → LLM classifies each one with structured JSON output → Aggregator deduplicates, calculates totals, identifies ghosts → FastAPI serves the dashboard.</p>
<p>Everything runs locally. No OpenAI API. No cloud processing. Your email content never leaves your machine. The LLM is on my hardware. The dashboard runs on a Linux box in my DMZ behind Caddy.</p>
<p>From idea to live product: 18 hours. That includes writing the scanner, building the classifier, creating the dashboard, deploying across nodes, and scanning my own email to generate real demo data.</p>
<h2>What It Found in My Email</h2>
<p>Nine subscriptions. $1,517.62 per month. Including AT&T wireless ($682/mo for family plan), Slack Pro ($85.33/mo for the federation workspace), Anthropic Claude ($20/mo), Plex Pass ($69.99/year), and Nate's own Substack ($200/year — not a ghost, that one's earned).</p>
<h2>The Confession</h2>
<p>This scanner runs on the same sovereign AI federation I've been building for 11 months. The same 7-specialist council that governs the federation classified the subscription emails. The same thermal memory architecture that holds 96,000+ memories served the results. The same infrastructure that processes council votes deployed the dashboard.</p>
<p>The Subscription Trimmer isn't a separate product. It's what happens when you have a production AI system and you point it at a real problem. 18 hours because the hard part was already built.</p>
<p>The scanner will be open source on GitHub. Your data stays on your machine. No accounts, no cloud, no third-party APIs. Sovereign.</p>
<p><a href="https://ganuda.us/trimmer">ganuda.us/trimmer</a> — see what you're paying for.</p>
<p><em>For Seven Generations.</em></p>