Greg Spero. Good meeting you tonight.
I'm a musician, and I run The Recording Club, a music studio club in Santa Monica. The system is called Gregbot. I talk to it from my phone, it works across 27 Claude Code sessions at once, and nothing it writes to anyone outside my team goes out until I press Send.
A server I talk to, agents that act, and a memory they share.
A Node server that drives a tmux grid of Claude Code sessions. I hold a button on my phone and talk. The words show up as I speak and land in whichever session I picked. Transcription runs on a local Whisper server. Each session conducts its own team of workers.
The agents and apps that do things: a CRM over my contacts and messages, life goals, the social video scheduler, the staff app and alerts for the club, and a send desk where every outgoing text and email waits for my click.
One fact per file, linked by name and loaded by hooks when a session starts. 241 of those files are my own rules and corrections. Behind them, a 4.9 GB SQLite database indexes contacts, messages, notes and recordings from a wearable. An Obsidian vault holds the rulebook for how Claude should work for me.
One Mac, my phone, and nine machines around the club.
Eight things from the last few months. Tap one for the detail.
I asked for it by voice. It shipped the same evening.
Open now.gregspero.com21 project cards, a filter that strips money, names and internal links before every publish, and all 13 links clicked in a hidden browser before I saw it.
A replacement for Sibelius, built around the keyboard.
Open notation.gregspero.comA caret editor with 103 commands, an MCP server so Claude can edit scores, an Electron desktop build, and MusicXML that MuseScore 4 opens. 261 unit tests and 9 end-to-end tests. The keystroke target isn't met yet: 129 keys against about 105 in Sibelius.
Members were cancelling without anyone telling me. Now a detector texts me about each one.
Every 15 minutes it reads Stripe (read-only), the club database, check-ins, iMessage and Gmail, then texts me a link to a retention board for that member. Two real alerts fired the afternoon it shipped.
If a contractor is scheduled over 9 hours, or works 9 with no break, my studio manager and I get a text.
The text opens the exact shift in the editor. It hooks every path that writes the schedule, plus a sweep every 5 minutes, with 16 tests and a kill switch. It shipped tonight, so no real alert has fired yet.
I talk to Grok while driving and it relays to Gregbot.
A webhook connector lets Grok ask and Gregbot answer. The responder is a sandboxed, read-only Opus run with deny rules on secrets, an output filter for key patterns, a 25-turn cap, rate limits and a kill switch. Today it carried 78 relays. One of the screens below is the page that sorted them.
Spirit Fingers Trio has a release on Nov 17. The site and brand kit were rebuilt in four parallel lanes.
Open spiritfingerstrio.comA v2.0 brand kit, 24 social templates and a new site, reviewed three rounds by a Fable judge until it cleared the 3 out of 4 bar.
Each of 65 artists and investors on Tiny Records gets a magic-link portal with their own statements.
A pipeline parses distributor statements, builds the ledger and reconciles the bank. 67 audit checks have to pass before every deploy.
My Gongyo meditation videos, rendered and scheduled while I slept.
18 long videos and 4 Shorts, mastered to loudness spec, uploaded private with weekly release dates from Nov 3 to Mar 2 inside the YouTube quota. Each upload was checked through the API.
When a rule written as an instruction didn't hold, it became code that runs before every tool call. 34 hooks, 18 of them guards.
Nothing goes to anyone outside my team until I click Send on that exact message. Approval is a one-shot token tied to a hash of the message, and Claude can't approve its own send. It went in after two texts reached real people during app testing.
since Jul 17, 2026No link goes out until system and public DNS agree, the page returns a real 200, and a hidden browser renders it. Built the same night a link I was sent showed a 404.
since Oct 14, 2026Fable sub-agents run only as planner or judge. Any agent spawned without an explicit model is blocked.
Claude's tools can't read the fleet and signing keys, fleet-wide commands need my logged OK, and the local grid API refuses unkeyed writes.
Native browser dialogs get cleared within a second, and automation can't navigate away from a half-filled form.
Captured tonight. Names, numbers and anything personal are blurred or cropped out.
