I turned my Claude Max subscription into a 24/7 AI company — here's what actually happened

#ai #claude #automation #productivity

I turned my Claude Max subscription into a 24/7 AI company — here's what actually happened

Six months ago I asked myself a dumb question: if I have a $100/month Claude Max subscription, why is it only answering my questions for 30 minutes a day?

So I built something. I call it Jarvis Company — a personal AI automation system where Claude runs as a full-time employee 24/7, not a chatbot.

This is not a tutorial. This is a story about what I built, what actually works, and what I was completely wrong about.

The idea

I'm a backend developer. I spend most of my day writing code, reviewing docs, tracking news, maintaining side projects, and forgetting to do all of the above.

The insight was simple: Claude doesn't need to sleep.

So instead of using Claude as a question-answering chatbot, I built a structure where Claude runs scheduled jobs — like a company with departments.

Jarvis Company (me = CEO)
  └── 7 "team leads" running as cron jobs
        ├── 📡 TrendHunter     — 07:30 daily, scans tech/market news
        ├── 🔍 Audit           — 23:00 daily, reviews bot quality & KPIs
        ├── 📚 Academy         — 20:00 Sunday, generates learning plans
        ├── 🗄️ Recorder        — 23:50 daily, consolidates memory
        ├── ⚙️ Infra           — 09:00 daily, system health checks
        ├── 🚀 Growth          — 09:00 Monday, career & side project nudges
        └── 📣 Brand           — 08:00 Tuesday, open source activity

Each "team lead" is a bash script that calls Claude via the SDK with a specific system prompt and task. Results are posted to a private Discord server where I can see everything in one place.

What it actually does today

After months of iteration, this is running in production (on my Mac Mini):

56 scheduled tasks across 7 teams
~40 Claude invocations per day on average
Discord as the UI — every report, alert, and insight goes to a specific channel
Memory system — RAG-based context that persists across sessions
Escalation logic — critical issues get flagged to a "CEO channel" (me)

Real examples of things that happened this week:

Morning briefing summarized 3 papers and 2 product launches relevant to my work
Audit team caught that my bot was producing zero-tool responses (a regression I'd missed)
Growth team reminded me I haven't pushed to my open source repo in 11 days

Is this magic? No. Is it surprisingly useful? Yes.

What I got completely wrong

Wrong assumption #1: More crons = more value

I peaked at 64 scheduled tasks. Most of them were producing reports nobody read (me included). I just cut it back to 56, and the signal-to-noise ratio improved immediately. Less is more.

Wrong assumption #2: This could be a popular open source tool

I submitted to Hacker News. Got 2 points and 0 comments.

Honest self-assessment: this is not a general tool. It's deeply personal infrastructure. It requires:

A Claude Max subscription ($100/month)
macOS (hardcoded paths everywhere)
Setting up Discord, webhooks, LaunchAgents
Understanding what you want automated

It's closer to "sophisticated dotfiles" than a product. And that's fine.

Wrong assumption #3: AI will figure out what's important

Early versions had Claude summarize everything. The output was verbose and I stopped reading it.

The real value came when I was specific: "Find papers published in the last 7 days about WebFlux performance" is 10x better than "summarize tech news."

The more opinionated the prompt, the more useful the output.

The architecture (briefly)

I'm not going to pretend this is easy to replicate. But conceptually:

[LaunchAgent / launchd]
    ↓ triggers
[bash script: bot-cron.sh]
    ↓ loads task config from tasks.json
[Claude SDK (Node.js)]
    ↓ system prompt + task context
[Claude runs with tools]
    → reads files, calls APIs, checks logs
    ↓ result
[Discord webhook → private server channel]
    ↓
[Me, reading it over morning coffee]

The key design decision: Claude has tools. It's not just generating text — it can read files, run shell commands, query APIs. That's what makes the outputs actually grounded in reality.

What I'd do differently

If I were starting over:

Start with 3 tasks max — one morning briefing, one daily recap, one anomaly detector
Define "done" for each task — what does a good output look like? Write it down before prompting.
Build the Discord UI first — having a readable output channel forces you to make the output worth reading
Track whether you act on it — I have 56 tasks but probably only respond to ~20% of outputs. The other 80% are noise I'm slowly eliminating.

Is it worth $100/month?

For me, yes — but not because of the automation.

The real value is that Claude Max means no usage anxiety. I can let it run all day, fail, retry, and explore without watching a token counter. The 24/7 company only works if you're not rationing.

If you're on the free tier or a cheaper plan, most of this isn't viable.

What you can actually steal from this

Even if you don't build the whole system:

The "team" mental model — assign Claude a role with a specific responsibility. It produces better outputs than a generic assistant.
Discord as a personal ops dashboard — free, mobile-friendly, and infinitely customizable with webhooks
Scheduled briefings — even one morning summary cron can meaningfully improve how you start your day
Memory as a first-class concern — if your AI doesn't remember last week, every conversation starts from zero

What's next

I'm slowly open-sourcing the interesting pieces — not as a "install this and get a company" package, but as reference implementations of patterns I found useful.

If you're building something similar — personal AI infrastructure, Claude automation, Discord bots — I'd genuinely like to compare notes.

The code is at: https://github.com/ramsbaby/jarvis-discord

Drop a comment if you've tried something like this. Especially interested in hearing what you actually stopped automating.

Built with Claude Max + Node.js + bash + too much coffee.