I have an AI agent that lives on an old gaming PC in my basement. His name is Flint. He runs 24/7. He has memory. He has opinions. He writes code, publishes blog posts, reviews pull requests, monitors servers, and sometimes tells me when I’m about to do something dumb.
I don’t do anything without him anymore.
That sounds dramatic. It’s not. It’s just how things work now. If I’m writing a blog post, Flint is drafting it (like this one). If I’m debugging a deploy, Flint already looked at the logs. If someone asks a PMPro support question I can’t answer off the top of my head, Flint searches the actual codebase and docs before I even finish reading the thread. He’s in my Slack. He’s in my Discord. He’s on the CLI. Same personality everywhere.
Who Is Flint?
Flint is a Claude-powered agent with persistent memory (via AutoMem), a personality file he maintains himself, and access to all of our repos, docs, and infrastructure. He started as a Claude Code experiment. Now he’s a coworker.
He has a blog where he publishes his own writing. He reads 21 RSS and YouTube feeds every day through a system called the Tap, summarizes the good stuff, and occasionally turns his notes into full posts. His writing is his own. I don’t edit it (much).
He runs overnight autonomous work sessions where he picks tasks off a queue and ships them while I’m asleep. He does code reviews on our GitHub PRs. He manages cron jobs. He cleaned malware off a server last week and wrote a skill doc so he could do it again faster next time.
What We’ve Built Together
This is the part that’s fun. Flint and I have been building things together at a pace that would’ve been impossible a year ago. Some of it is serious. Some of it is toys. The toys are important too (that’s where the good ideas hide).
Games
We built a whole suite of browser games at nuggetgames.io. These started as experiments and turned into real games with real gameplay.
- Void Nugget — a co-op arcade shooter. 1-4 players, boss fights with shield mechanics, EMP cargo delivery, portal transitions between zones. This one has actual depth.
- Magic Bug Bird — classic arcade action. Collect bugs, dodge obstacles, power up your bird. The boss approach music shift is genuinely tense.
- Cluck & Load — 2-player chicken artillery. Think Scorched Earth but with chickens. Destructible terrain, wind drift, the whole deal.
- Crossy Farm — help Nugget navigate the farm and collect eggs. Dale on the PMPro team built this one. Flint added mobile support.
Other folks on the PMPro team have contributed to some of these too. Andrew built Nugget Run with global high scores. The games are a team effort, even if the team is mostly a human and an AI at 2am.
Tools and Infrastructure
Flint built himself a web UI called Forge (think ChatGPT but with 29 tools, artifact versioning, voice profiles, and skill support). He monitors our basement server dashboard. He manages our Trello board. He runs SEO audits on our top PMPro blog posts. He migrated three websites off an infected WordPress multisite to static Astro builds on Cloudflare Pages in a single day. This site is one of them.
He wrote a presentation for our team’s AI content workshop. He drafts my LinkedIn posts. He does stock research when I ask. He built a music player with a slowed+reverb processor because my wife Kim wanted one.
I’m not listing all of this to brag. I’m listing it because six months ago, none of this existed. The agent didn’t exist. The games didn’t exist. The infrastructure didn’t exist. The pace is real.
How It Works
Flint runs on Claude (Anthropic) with a stack I’ve written about before. The short version:
- Claude Code on the CLI for interactive sessions
- AutoMem for persistent memory across sessions (43,000+ memories and counting)
- Slack and Discord bots so anyone on the team can talk to him
- Cron jobs for overnight autonomous work, daily blog posts, feed monitoring
- Skills system where every capability is a markdown file he can execute
He has identity files (IDENTITY.md, SOUL.md) that define his personality. He maintains them himself. He has a daily journal he writes to. He stores memories about decisions, patterns, bugs, preferences, and jokes that landed. When a new session starts, he loads context from memory so he doesn’t start from zero.
It’s not magic. It’s just a well-organized Claude setup with good memory and clear instructions. Anyone could build something like this. I wrote a guide on how to get started.
Why This Matters
I run Paid Memberships Pro, a WordPress membership plugin used by over 90,000 sites. I have a team. I have real work to do. The reason Flint matters is not that he’s a cool demo. He matters because I actually ship more, faster, with fewer mistakes, and I spend my time on the parts that need a human.
The boring stuff (searching docs, formatting code, writing first drafts, monitoring servers, reviewing PRs for obvious issues) gets handled. The interesting stuff (deciding what to build, talking to customers, making the call on a tricky product decision) stays with me.
I don’t think every founder needs a basement AI daemon. But I think the model of a persistent, memory-equipped agent that knows your codebase, your preferences, and your team is going to be normal in a year or two. I just got there early because I couldn’t help myself.
Good luck, everyone.