Most "I built an AI agent" posts come with a screenshot and a vibe. This page is the receipts — live counts from my running stack, refreshed each morning at 06:15 ET. No render-on-load tricks. The system actually does this work.
Cost of running
frontier-model cost · lifetime
Every frontier-model call NEXUS makes on its own (2,690 of them), at API list prices, since logging began 2026-05-24. Routine work runs on local models and isn’t in this number. Supervised coding sessions are on a flat subscription and aren’t either.
frontier-model cost · last 30 days
Most of it is quality checks: a frontier model reviewing NEXUS’s own output before it’s used.
posts shipped · last 7 days
Drafted by NEXUS, reviewed by me, fired on weekly slots. Receipts below.
Last 7 days
Agents
Autonomous workers running in my stack — each with a single job, on its own schedule, self-healing, around the clock.
Brain 3
Runs the cycle every minute, sends each incoming signal to the right worker, and drafts new capabilities that only ship after review.
Want this in your business? →Content 4
Turns notes and research into drafts. I approve every post before it publishes on schedule.
Want this in your business? →Research 1
Gathers the overnight signals worth reading in the morning.
Want this in your business? →Opportunities 1
Scans creator channels for ideas worth turning into offers.
Want this in your business? →Voice 1
Reads the morning brief aloud while I get ready.
Want this in your business? →Memory 2
Condenses what happened into long-term memory, and notes what worked and what to change.
Want this in your business? →Recently shipped
The most recent posts NEXUS drafted, I reviewed, and the stack published.
You're probably wondering how I make these videos. I prop my phone up in my room, act the scene myself, and AI turns me into whoever the scene needs. No studio, no actors. People keep asking how I keep up with posts ...
Read on linkedin →Honestly the fix for AI design slop is a CSS library that tells the model what good looks like - that's the whole problem in a repo. 34k stars on a band-aid. The stars are real, the utility might be real. But you're ...
Read on twitter →I acted a shot once in my room, phone propped up, and decided afterward who was in it. Left is me. Middle is what the AI actually sees. Right is what it makes. The middle panel is the whole trick. It's a depth map: ...
Read on linkedin →Radisson's ChatGPT plugin converts 1.5x organic in 6 weeks. The plugin didn't earn that. The 7 years rebuilding their direct booking channel did. Own the checkout, rent the discovery. Every team I've fixed this for ha...
Read on bluesky →Interfaces that render before the model finishes thinking feel instant. Instant is not the same as correct. A one-word rename once cost me hours of work, and nothing flagged it until I went looking. What has your tool...
Read on bluesky →This page is a static snapshot, pushed once each morning by the same stack that runs everything else. No backend · no live API · no telemetry from your browser. The essays tell the story behind the numbers.