GoodStrataYour strata manager is now an open-source AI agent. It ran our building from a MacBook.
Every apartment owner in Australia pays a strata manager — typically $300–$600 per lot, per year — for work that is mostly administration: issue the levies, chase the arrears, book the plumber, send the AGM notice, take the minutes. The work matters enormously and is done indifferently, because the incentives are wrong: managers earn commissions on the insurance they place and the contractors they choose, and the owners footing the bill can't see any of it.
GoodStrata is our answer: an open-source platform where AI agents do all of that work, and the humans in the building do the only thing that should ever have been theirs to do — decide. Approve the budget. Approve the quote. Vote at the AGM. Everything else runs itself, and every single action lands on an append-only event log you can audit down to the individual model tool-call.
It's free for owners corporations. Not freemium — free. The platform earns a payment-processing margin when levies move and a commission from contractors who win work through it. No subscriptions, no lock-in — and because it's Apache-2.0 open source, "no lock-in" is a checkable claim: self-host it with nothing but PostgreSQL and one Node process.
We ran a whole building on a laptop, with no cloud AI
The obvious objection to an AI strata manager is trust. So we built for the most paranoid deployment first: everything local. To prove it, we seeded a realistic building — 48 Rose St, Fitzroy, a 12-lot walk-up with an adopted $60k budget, quarterly levies issued, 11 owners paid up, one owner 47 days in arrears, and a fresh "water stain on the lot 9 ceiling" report — and pointed the agents at Qwen3-30B-A3B running in Ollama on a single MacBook (M4 Max). No API key. No data leaving the machine.
Here's the timelapse of what happened when the platform booted.
T+0:00 — the dispatcher wakes the agents
The event bus (plain Postgres — an append-only log with LISTEN/NOTIFY) replays anything the agents haven't seen. Three of them go to work simultaneously.
T+1:38 — the meetings agent drafts minutes, and refuses to make things up
Last year's AGM record had no attendance entered. The drafted minutes:
Quorum: Not achieved (0/130 entitlements represented).
Business transacted: None. The meeting was adjourned due to lack of quorum.
A 30B local model, given the chance to invent attendees and motions, wrote "business transacted: none" — because the structured record it was handed said so, and the prompt says minutes record facts. This is the design everywhere: the model narrates; the ledger is the truth.
T+2:21 — the finance agent chases the arrears, with figures it cannot get wrong
Lot 7 is 47 days overdue — stage 3 of the statutory-style ladder that code, not the model, walks daily. The agent's job is the words:
Final Notice: Outstanding Levies – Lot 7
Dear Pat Latimer, your levy payments for Lot 7 are $1,153.85 overdue (47 days), plus $14.86 penalty interest. This is a final notice; please contact us to arrange payment or discuss a hardship plan…
Those figures are correct to the cent — and they'd be correct even if the model hallucinated, because GoodStrata appends a code-generated statement block to every money email. The LLM never computes money. Levy apportionment (largest-remainder, sums exactly), penalty interest (actual/365), payment matching, vote weighting: deterministic engines with exhaustive tests.
T+9:32 — the maintenance agent triages the leak and hires the roofer
"Water stain caused by roof leakage (common property) beneath roofline, worsening with rain; not internal fittings or owner's lot issue." — category: roofing, urgency: high, common property: yes
That's the exact legal distinction — owners corporation vs lot owner responsibility — that fills strata Facebook groups with grief, made correctly by a local model. It then proposed Rapid Roofing with a $350 inspect-and-repair scope. $350 is under the scheme's auto-approve threshold, so the platform — never the model — dispatched it immediately and emailed the contractor a bounded work order: "Approved amount: $350.00 (do not exceed without written approval)." The model even tried the same tool twice; the state guard rejected the duplicate. One work order exists.
Day 61 — the machine stops and asks the humans
We fast-forwarded lot 7 to 61 days overdue. The daily sweep crossed the day-60 line, and the finance agent did the one thing it's not allowed to do alone — it opened a committee decision:
Commence debt recovery — lot in arrears 61 days
Lot 7 (Pat Latimer) 61 days overdue with $1,153.85 levies and $19.28 penalty interest. Recommend committee approval to commence recovery.
The agent's run ends there, in a state literally called awaiting_decision. A committee member opened the decisions inbox and tapped Approve — the single human act in this entire story. A code executor (not the model) then sent the formal demand: total now payable $1,173.13, 14 days, hardship option stated.
The payment closes the loop by itself
Pat paid by the unique payment reference on the notice. The webhook was signature-verified, matched to the notice, posted to the lot ledger, split across the admin and maintenance funds to the cent, receipted, and emailed — status: paid. Nobody reconciled anything.
The scoreboard
| Events on the audit log | 110 |
| Agent runs | 4 (maintenance, finance ×2, meetings) |
| Total model tokens | ~21,000 |
| Emails drafted & sent | 4 — final notice, work order, formal demand, receipt |
| Work orders dispatched | 1 ($350, threshold-routed by code) |
| Human decisions required | 1 (approve debt recovery) |
| Money computed by an LLM | $0.00 |
| Hardware | one MacBook, Qwen3-30B-A3B via Ollama |
Around 21k tokens of local inference ran a month's worth of strata management. At hosted-model prices that's a few cents; on your own machine it's electricity. This is why GoodStrata can be free.
Why you can trust it (the honest version)
Not "trust the AI" — trust the architecture:
- Append-only event log, enforced by a database trigger. Every action carries its actor (user, agent, or system) and a causation chain back to what triggered it. The Agents tab shows every model tool-call.
- Decision gates. Budgets, over-threshold works, debt recovery, rule matters — agents can only propose, with evidence attached. Approval executes exactly the proposed action, via code.
- Statute as code. The 14-day meeting notice rule, entitlement-weighted voting, 75% special resolutions, the s 89 bar on arrears lots voting, the arrears ladder — unit-tested functions, not prompt text.
- Deterministic money. If every model on earth were replaced with a random number generator tomorrow, your levies would still sum to the budget exactly.
Get it
- Self-host:
docker compose --profile app up— Postgres plus one container. PointAI_PROVIDER=localat Ollama, oranthropicfor hosted Claude. Demo building included:pnpm seed:demo. - Managed: goodstrata.com.au — same open code, hosted, with real NPP/PayID payments. Coming soon to Australian buildings.
- Source: Apache-2.0, on GitHub. The e2e test drives the entire story above — onboarding to AGM — in a real browser on every commit.
Your building's money, minutes, and maintenance — run perfectly, audited completely, and owned by no one but you.
Artifacts from the simulation in this post (agent transcripts, the event log, every email) live in the repo under docs/blog/sim-artifacts/ — regenerate them yourself with pnpm seed:demo and a local model.