FocusNow AI Club, meet-up #4 – 29 July 2026.
The fourth AI Club session was a live walkthrough of an AI environment running on one machine at home – what it takes to build, what it can do, and where it stops making sense.
The setup
The hardware is modest: a Mac Mini M4 with 16GB of unified memory, accessed remotely over screen sharing. It draws around 70 watts, so it runs permanently at negligible cost. Two agents run side by side, with local tools letting them process images and scanned PDFs, not just text.
The most interesting component was the memory layer. Assistants lose context once a conversation exceeds the model’s window, so both agents read and write to a shared local knowledge vault. Anything worth keeping – how an application was built, which credentials belong where – is still there weeks later.
Day-to-day interaction happens through WhatsApp. No terminal, no dashboard.
What it enables
Most capabilities came from publicly available agent skills rather than custom code:
- A full expense-tracking web application, built without its owner writing a line of code. Bank statements drop in as CSV, get categorised automatically, and feed sortable reports.
- Email, calendar, file storage and reminders, connected directly to the agent.
- Timesheet capture over chat, logging work as it happens for month-end submission.
- Premises monitoring – natural-language queries against local CCTV, no footage leaving the property.
None of it required deep engineering.
The ServiceNow angle
For platform work, the significant development is that ServiceNow now publishes its documentation as clean Markdown on GitHub with an llms.txt index built for LLM consumption, refreshed monthly. Previously the docs site blocked crawlers – a large part of why AI tools so confidently invented ServiceNow classes that do not exist. Indexed locally, questions resolve against authoritative documentation.
Off-the-shelf skills also cover ITSM, CMDB and the Table API, making plain-language portfolio reporting possible without custom development. An attempt to go further – requirements to user stories to working configuration in a PDI – was abandoned because the code was poor, and is now being rebuilt.
Local or cloud
Local AI is hard to justify on cost alone. Serious models need 64GB or more, and machines at that spec approach €5,000, while a month of cloud usage came in under $10. The real case is data residency: a fully offline setup sends nothing anywhere, a genuine constraint in regulated sectors.
Key takeaways
- A capable local setup is achievable on modest hardware, if expectations match the memory.
- Most capability comes from existing agent skills, not custom development.
- A persistent memory layer is what makes an assistant useful over time.
- Choose local hardware for data residency, not savings.
Join the next one
The FocusNow AI Club meets monthly, with working demonstrations rather than slideware. If you are weighing where AI fits in your environment, get in touch.