BackIn Progress · Designing

MyMind

A private AI assistant running entirely on a Mac mini in my home. Six connected agents that know everything about my life — and never let it leave the network.

Why I'm building this

The most valuable AI assistant is the one that knows you best — but that's also the most dangerous one to put in the cloud.

Every cloud AI tool asks for your data. Your health records, your financial accounts, your documents, your habits. The deal is simple: hand it over and we'll give you something useful back. But the data never really comes back — it lives on someone else's servers, gets used to train models, becomes a target for breaches, and shapes products you'll never see.

MyMind is the opposite bet. A Mac mini sitting in my home runs local LLMs and a network of agents that know my health, finances, travel, and life admin. None of that data ever touches a cloud API. The model runs on my hardware. The data sits on my disk. The only thing that leaves the house is a curated query, when I choose to send one.

It's also a forcing function for me as a builder. Cloud APIs make AI feel like magic. Running it locally forces you to make real decisions about model size, latency, memory, and orchestration — the kind of decisions you never have to think about when you're just calling Claude.

The Agents

Six agents. One brain.

Six specialized agents, all reading and writing to a shared local database, all powered by a single LLM running on the Mac mini. Each one owns a domain and shares context with the others.

Health Hub

Health

All your health data in one place — without it ever leaving your home.

Apple Health data, doctor visits, dental records, lab results — all stored locally and unified into a single timeline. The agent flags anomalies, summarizes trends, and prepares you for upcoming appointments.

What it surfaces, unprompted

›You haven't been to the dentist in 6 months and I don't see anything on your calendar — want me to book?

›Your resting heart rate has been climbing for 3 weeks. Worth mentioning at your physical.

›Your last cholesterol panel was 12 months ago. Most providers recommend yearly.

›You slept under 6 hours four nights this week. Worst week since February.

Design decisions

(4)

Daily Stock Researcher

Finance

A morning briefing on every stock you care about — synthesized overnight.

While you sleep, the agent scans news, SEC filings, and insider trades for every ticker on your watchlist. By morning, you have a 3-5 bullet summary per stock with a green/yellow/red signal.

What it surfaces, unprompted

›NVDA dropped 4% overnight on China export news. Three of your other holdings are exposed.

›An insider sold $40M of META yesterday. First time this year.

›Your portfolio is up 2.1% pre-market. Tech is leading.

›You added AMD to your watchlist 3 weeks ago. It's up 18% since.

Life Admin Assistant

Life Admin

Every important document at your fingertips, with reminders you actually want.

Drop your passport, insurance policies, warranties, and contracts into a folder. The agent reads them, tracks expiry dates, and alerts you 90/60/30 days out. Ask it anything and it answers from your actual documents.

What it surfaces, unprompted

›Your passport expires in 87 days. Renewals are taking 8-11 weeks right now.

›Your car registration is due next month. Last year it was $187.

›You're paying for Dropbox and iCloud. Want to consolidate?

›Your renter's insurance auto-renews in 14 days at a 12% higher rate.

Second Brain

Knowledge

Everything you've ever saved, finally findable.

Auto-ingests bookmarks, Kindle highlights, saved articles, PDFs, and notes. No behavior change — you just bookmark like normal. The agent makes everything searchable in plain English and surfaces relevant items across other agents.

What it surfaces, unprompted

›You bookmarked 4 articles about local LLMs this week. Want a synthesis?

›I noticed you saved a Cursor tip last month — relevant to what you're building today.

›Your Kindle highlights from Hard Things keep matching topics you're thinking about lately.

›You've saved 12 woodworking joinery tutorials over the past year. Want a summary?

Personal Finance

Finance

A net worth dashboard that updates itself, with anomaly detection built in.

Plaid connects every bank, credit card, and brokerage account. The agent tracks net worth daily, flags unusual spending, monitors points balances, and feeds clean financial data into the Stock Researcher and Travel Agent.

What it surfaces, unprompted

›You spent $340 on food delivery last week — 2x your usual.

›Net worth crossed a new high yesterday.

›Chase Sapphire annual fee hits in 9 days. Was the value worth it this year?

›You're 60% to your savings goal for the quarter.

Travel Agent

Travel

Knows what you've got and finds the best way to use it.

Tracks every loyalty program — credit card points through Plaid, airline and hotel miles through browser automation. When you plan a trip, it suggests the best redemption, flags transfer bonuses, and checks your passport expiry against the destination requirements.

What it surfaces, unprompted

›Chase to Hyatt transfer bonus just opened — 30% extra through April 30.

›Your United Premier status expires in 82 days. You're 4 segments short.

›Lisbon flights for May just dropped 22%. You bookmarked this destination in February.

›Your passport expires before that Brazil trip. Renewal won't make it in time.

How they talk to each other

Health + Life Admin → dentist due alertsFinance + Travel → points redemptionsLife Admin + Travel → passport expiryStock Researcher + Finance → portfolio viewSecond Brain → surfaces across all agents

Design Philosophy

Every feature has a build-vs-buy-vs-skip decision.

Local first, cloud never

Every component runs on the Mac mini — LLM inference, document storage, agent orchestration, scheduled jobs. The only network calls are reads (Plaid for finance, Playwright into account portals for points). Nothing leaves the house.

Manual when manual is reliable

The temptation with AI projects is to automate everything. But auto-login to patient portals, OAuth flows that break monthly, scrapers that fight CAPTCHAs — these are the things that turn a passion project into a maintenance nightmare. Manual file drops and shortcut automations are boring, but they work forever.

Boring infrastructure for the unboring stuff

The agents are AI-native and ambitious. The plumbing is intentionally boring — Postgres, Express, cron jobs, Plaid SDK, Playwright. Spend the complexity budget where it actually matters. Save the magic for the model.

Connected, not siloed

The interesting product moments come from agents talking to each other. Health + Life Admin = dentist due alerts. Finance + Travel = best points redemption. Each agent is useful alone, but the system gets exponentially more valuable as agents share context.

Stack

Built on consumer hardware.

Everything runs on a single Mac mini in my home. The whole stack is intentionally boring — the only ambitious part is the agents themselves.

Hardware & Runtime

Agents & Orchestration

Data Ingestion

Roadmap

The plan from now to fully connected.

I'm sharing this early on purpose. Most portfolio projects only show finished work. I'd rather show the thinking, the tradeoffs, and the decisions before the code exists — that's the part that actually matters in product work.

Phase 0

Foundation

In progress

Get the hardware running and prove I can hold a useful conversation with a fully local model.

Mac mini ordered, OS configured for headless 24/7 operation
Ollama installed, Llama 3.2 running locally with measurable latency
Postgres set up as the single shared data layer for all agents
First end-to-end test: a query that uses local LLM + local data

Phase 1

First agent — Health Hub

Prove the full loop on the highest-stakes use case. If health works, every other agent is the same pattern.

Apple Shortcuts → nightly HealthKit sync to Mac mini
Watched folder for doctor & dental PDFs with auto-extraction
Health timeline UI — see trends across labs, vitals, visits
First useful output: pre-appointment summaries and anomaly flags

Phase 2

Money agents

Planned

Plug in Plaid once and unlock two agents that share the same data foundation.

Plaid connection for bank, credit cards, and investments
Personal Finance agent — net worth, spending, anomalies
Stock Researcher — nightly cron, watchlist scans, morning briefing
First cross-agent moment: portfolio-aware stock recommendations

Phase 3

Life infrastructure

Planned

The boring-but-essential agents that turn MyMind from a tool into a daily habit.

Life Admin Assistant — passport, insurance, warranties, expiry alerts
Second Brain — bookmarks, Kindle highlights, PDFs, plain-English search
Recurring task and subscription tracking
Cross-agent test: dentist due alerts from Health + Life Admin

Phase 4

Travel + the full network

Planned

The final agent and the moment all six start talking to each other in real ways.

Travel Agent with Playwright-scraped points & miles balances
Best redemption suggestions tied to Personal Finance and trip planning
Passport expiry from Life Admin → travel readiness checks
Second Brain surfacing relevant saves across every other agent

Phase 5

Build in public

Planned

Once it's real, share what worked, what broke, and what I'd do differently.

Write up architecture decisions and benchmarks
Open source the orchestration layer if it's reusable
Publish a teardown of running real agents on consumer hardware

Want to follow along?

MyMind is being built in public. Reach out if you want to compare notes on local AI, agentic workflows, or just hear how this is going.

Get in touch Back to all work