All case studies
Operating platform

Stealth client

We built an operating platform for an Australian property development business: boards, documents, dashboards and files in the browser, with a built-in assistant and an MCP connection for the AI clients the team already uses.

Client name withheld at their request.

Stealth client

Engagement focus

Project workspace, governed AI, MCP and Microsoft 365

Category
Operating platform
Status
Live in production
Engagement
Project workspace, governed AI, MCP and Microsoft 365

What we built

  1. 01

    A browser workspace and built-in assistant, with the same governed tools available to external Claude and ChatGPT clients through MCP.

  2. 02

    Claude powers the built-in assistant; identity, permissions, approvals and audit stay in the platform rather than in the model.

  3. 03

    SharePoint retrieval revalidates every result against live Microsoft permissions before returning it, and fails closed when access is unclear.

Outcome 01
Projects, documents, approvals and files in one place instead of across spreadsheets and inboxes.
Outcome 02
AI that works on live company data, with no downloading, re-explaining or manual filing.
Outcome 03
A custom platform the client owns, with external AI clients and built-in model usage kept as distinct operating costs.
Operational auditCustom platform buildMicrosoft 365 integrationMCP tool layerAI team training

How it's built

Workspace and data
Next.js and React deliver the browser workspace. Supabase stores workspace data and media; server-side access checks govern boards, documents, dashboards and sharing.
Microsoft 365 and retrieval
Delegated Microsoft Graph access connects SharePoint, OneDrive, Outlook, calendars and Teams. PostgreSQL lexical search, Graph search and activity-aware ranking find sources; results are checked against the signed-in user's live Microsoft permissions before being exposed.
AI and approvals
Claude powers the built-in assistant through the Vercel AI SDK. OAuth-protected MCP exposes the same governed tools to external clients. AI Canvas designs use esbuild, Recharts, D3 and Three.js inside a sandbox before approval; OpenAI voice transcription is isolated from business actions.
Documents and hosting
PDF.js and pdf-lib support PDF workflows; ExcelJS, docx and PptxGenJS support Office outputs. Vercel hosts the application, API routes and scheduled jobs; Resend handles notification email.

Technology used

Web app

  • Next.js
  • React

AI

  • Claude
  • OpenAI

Database

  • Supabase
  • PostgreSQL

Integrations

  • Microsoft 365
  • MCP

Languages

  • TypeScript

Hosting

  • Vercel

The full story

One workspace for the work, and a way to point their own AI at it.

The brief was to give a property development team's work a home. Projects ran across spreadsheets, inboxes and manual approvals, with useful knowledge held by individual staff rather than shared through a system.

The less obvious half came out of the audit. The team already had AI. What they did not have was any way for it to touch real work. Every task became the same round trip: download the files, upload them into a chat, explain the project from scratch, get an output, then file it back by hand. The assistant was capable and completely disconnected.

The platform brings the browser workspace, its built-in assistant and external MCP clients onto one backend. Work holds boards, items, typed fields, versioned documents, dashboards, files, sharing, activity and automations. Claude or ChatGPT can reach the same workspace and Microsoft 365 through an authenticated MCP endpoint.

The built-in Work, dashboard and Canvas assistants use Claude through the Vercel AI SDK. External MCP clients bring their own reasoning model. Both paths use the platform's identity, permission-filtered retrieval, tools, approvals and audit trail. OpenAI is used separately for voice transcription; that route receives audio and cannot operate business tools.

Retrieval is deliberately unfashionable. There is no vector database and no embedding pipeline. SharePoint discovery runs on a PostgreSQL lexical index over file metadata and the first slice of extracted text, combined with Graph search and deterministic ranking by project, document family, revision and date. Every candidate that could leak a filename, path or even the existence of a file is revalidated against live Graph for that user before it is returned, and an indeterminate result fails closed.

Browser and MCP paths converge on the same server actions and the same access checks, so an edit made by an assistant is not a second, weaker implementation of the product. Scope was held just as firmly: Microsoft 365 remains the source of truth for its files, mail and calendar; the platform adds the operating workspace and governed actions around it.

Client reference

Testimonial available on request, subject to client approval.

Request a testimonial
Get started01

Tell us where your time goes.

We map where your week actually goes, then tell you what is worth building.

  1. 1We review your business context
  2. 2We book a short call or site visit
  3. 3You receive a systems map

Start with three details

No long questionnaire. We can fill the gaps on the call.

A direct conversation with the people who will build it.