We connect AI to the systems your business actually runs on.
Not another chatbot. Over the past few months we've built and hosted a fleet of Model Context Protocol servers — the secure bridge that lets an assistant like Claude read, draft and operate real systems: your phones, your mailbox, your files.
Model Context Protocol (MCP) is the emerging open standard for giving AI assistants real tools and data — not just conversation. An MCP server is the secure, permissioned bridge between an assistant and a system you already run. Ask a question in plain English; it does the work in the actual software and hands back the answer.
We design, build, harden and host these ourselves. That means the assistant can genuinely act — book the meeting, pull the call stats, draft the reply — while your data stays on infrastructure you control instead of being shipped off to someone else's SaaS.
The servers we've built
Each one is a real, running integration3CX phone system
Reads and safely operates a 3CX PBX — extensions, queues, call history with AI summaries and sentiment, inbound/outbound routing and business-hours profiles. Every write is preview-then-confirm with an audit trail. We run it against our own phones.
Microsoft 365
Full delegated Microsoft Graph access for a signed-in user — mail, calendar, Teams, SharePoint and OneDrive, contacts and To Do — plus a generic Graph escape hatch for everything else.
Google Workspace
Gmail, Drive, Docs, Sheets, Slides, Forms, Calendar and Chat under one AI connector — read, draft, edit and automate across the whole account without copy-paste.
Transcription
Turns call recordings, meetings and video into searchable, summarised transcripts an assistant can actually reason over — not just a wall of text.
File drop
Tokenised, expiring file transfer between people and AI assistants — move documents in and out of a workflow without email attachments or a shared drive left open.
Notebook automation
Drives a research notebook end to end — feed it sources, generate and pull summaries — as one step inside a larger automated workflow.
Bespoke line-of-business
Where a client's work lives in a specific system — a case manager, a job board, a booking platform — we build an MCP around it so their team can drive it by asking, not clicking. Scoped and hardened to that system.
How we build them
The part most "AI integrations" skipAn assistant with the keys to your mailbox and phone system is only as good as the fence around it. This is where the real work is, and it's the same standard on every server we ship.
- Self-hosted. Runs on infrastructure you control — your tenant data and call records never leave for a third-party service.
- Self-contained OAuth 2.1. Nothing is reachable without a key; each connection is approved explicitly, and access can be revoked.
- Breakglass recovery. A separate emergency key can wipe a server's stored credentials for a clean start if anything is ever compromised.
- Hardened by default. Non-root containers, firewalled to the reverse proxy only, rate-limited against scanning, TLS everywhere.
- Guarded writes. Anything that changes a system is preview-then-confirm, with an audit trail — no surprise edits.
- Packaged for handover. A client can stand one up from a single image and a web setup page, with no secrets baked in.
We build with AI, not just about it
The same tools we put in front of clients, we use ourselves. Deployments are driven and checked by AI agents working in parallel — and nothing goes live on a hunch. A bonus on the security side: our organisation is approved for dual-use cybersecurity use of Claude, so the pentesting and red-team work that's blocked by default elsewhere is on the table for us.
Adversarial review, every launch
- Independent AI agents attack each build across security, SEO and accessibility
- Every finding is verified before it's actioned — no false alarms acted on blind
- This very site was put through that review before it went up
AI where it earns its keep
- After-hours call triage and overflow on the phone system, escalating when it should
- Repetitive keyboard work — re-keying between systems — handed to a machine
- Summaries, transcripts and sentiment surfaced from calls and meetings automatically
If a task is "read from this screen, decide, then do that in another system," there's usually an assistant-shaped answer to it now — and we build the safe version.
We also do this for fun
The homelab is the R&D labHonestly? None of it started as a business plan — it started as a homelab that got out of hand. The techniques we sell earn their stripes on our own gear first: if a self-hosted AI server can't survive our network, it doesn't go anywhere near yours.
The lab it's tested in
- A multi-node Proxmox cluster running a stack of self-hosted services behind our own firewall
- The very MCP servers we put in front of clients live here first
- Monitoring, backups and alerting across the whole lot — that's the fun part too
We scan our own house first
- A hardened opnsense firewall at the edge — segmented, rules tightened, API-managed and audited
- A Greenbone / OpenVAS vulnerability scanner running on Kali Linux, sweeping the whole network on a schedule
- We find our own holes before anyone else does — the same discipline behind an Essential Eight uplift
The house runs itself
- Local-first Home Assistant across lights, climate, blinds, cameras and alarms
- An AI assistant wired straight in — "turn the studio off and arm the house" is a sentence, not a dashboard
- Keeps working when the internet doesn't, because it all runs at home
AI that earns a grin
- Cameras that can tell a person from a possum before they trip an alert
- A weekly automation that kills one household annoyance at a time
- We run a little community around it — Angry Dad Automates
The point isn't the gadgets. It's that everything we recommend — local-first, self-hosted, properly locked down — we live with ourselves, at home, every day.
When there's no API, you find a window
We wanted the family shopping list to land in the Woolworths cart on its own. One snag: Woolworths has no shopping-list API, and the site is wrapped in serious bot and anti-automation protection — built to stop exactly this kind of thing.
- So we did it the scrappy way — a headless Chrome browser driving the real website like a person, with a bot-blocker bypass running interference, syncing a Keep list straight into Woolies
- It's held together with sticky tape and breaks the day Woolworths moves a button — absolutely not enterprise-grade, and we'd never ship it to a client
- But that's the point: when there's no clean path to a system, we'll still find a way to a result — the same stubbornness we bring to real work, pointed at the grocery run
The hobby that shipped as a real product
ADHDTasker is a family chores-and-rewards app we build and run — the whiteboard on the wall, turned into a shared app. Built on Firebase, live as a web app today and in active beta on Google Play.
- Set up by just talking. A conversational "Guided Setup" builds the kids, sign-ins, routines, chores and rewards from a chat you type or say out loud — no forms.
- Yuck × time scoring. Points scale by how gross a job is and how long it takes, so the worst chores finally get volunteers.
- Real-time everywhere. The board syncs live across every device, with parent approvals, a points bank and rewards kids actually fight over.
- A real two-way Home Assistant integration — a proper HACS custom component, not a webhook. Finish a chore and the TV announces it while the room lights turn blue.
- Home Assistant reads the board too — per-kid banks, approvals and a leaderboard as sensors — and can create tasks straight back through the app's inbound API.
- The same product engineering — auth, real-time sync, device integrations — we bring to client work, shipped end to end on our own.
Got a system you wish you could just ask?
Whether it's your phones, your Microsoft 365, your Google Workspace or a line-of-business app nobody enjoys using — if it has an API, there's a good chance we can put a safe AI front door on it.