- PRIVATE
- STAGING VERIFIED · 7 OCT 2026
- FIRST RELEASE DEPLOYED · NOT YET LAUNCHED
Motera OS
An operating system for a small digital studio, built as my own independent project: CRM, delivery, a client portal, website production — and an AI layer that can draft and ask, but never act alone.
Interface rendered in Motera OS's own design tokens · demo data, not a screenshot
01 — The problem
A one-person studio still runs a whole company: enquiries, proposals, delivery, client updates, websites.
That work usually lives in five tools that don't know about each other. Meanwhile every new AI tool wants to send, post and deploy on its own. I wanted one system where work is recorded once — and where AI does real work without ever being trusted blindly.
02 — What's inside
Fifty-eight data models in one schema, organised into eight areas. Every area enforces tenancy and roles at the service and page level.
CRM
Companies, contacts, leads, opportunities and a pipeline board, with ownership and history.
BUILTCommercial
Proposals with versions, withdrawal and a PDF; competing proposals on one deal.
BUILTDelivery
Projects, tickets, tasks and “My work”, with business-day overdue rules.
BUILTClient portal
A separate session that shows a client only what was shared with them.
BUILT · READ-ONLYStudio & websites
Sites, versions, deployments, domains and audits, tracked per project.
BUILT · VIEWSAdmin centre
Members, roles, invitations, enquiry routing, intake channels and an audit log in plain words.
BUILTReporting
Sales, delivery and team views whose figures are checked against direct database queries.
BUILTIntelligence
Seventeen defined agent roles with runs, approvals and budgets. Seven work today, at least in part; ten are planned.
PARTLY BUILT
03 — One enquiry, end to end
A consultation request on the Motera website enters Motera OS and moves through the business without anything leaving it unapproved.
- 01Website enquiryIntake receipt; duplicates return the same reference
- 02AssessedLead Intelligence · Level 1
- 03Reply draftedCommunications · Level 2
- 04A person approvesThe exact email · Level 3
- 05Opportunity → proposalVersions, acceptance
- 06ProjectTasks, tickets, portal
04 — Decisions that shaped it
Agents can ask. They can't act.
Four autonomy levels. Reading and drafting run freely; email, outreach, paid model use and remote coding need a person to approve the exact payload. Production deploys need the Owner and have no executor at all. Billing, credentials and destructive changes have no agent path.
Budgets are held before the call.
Each model call reserves its worst-case cost atomically, so two runs at once can't overspend a budget that only one could afford.
The client portal is tested like an attack.
Guessed ids, revoked grants, another company's proposal on the same deal, the same person in another organisation — each is a test that must fail to see data.
Tests have to catch real defects.
New guards were broken on purpose to prove a test fails — by the time the portal shipped, 42 reintroduced defects had each been caught. A passing suite that can't catch a bug isn't evidence.
05 — Verification ledger
| Check | Result | Source |
|---|---|---|
| Automated tests | 977 passing, 0 failed, 0 skipped | Independent offline re-run · 25 Oct 2026 · in-memory database, no network |
| Interface sweep | 99 pages at 1366 and 375 px; no serious or critical axe issues | Automated interface sweep · 10 Oct 2026 |
| End to end | Enquiry → project 15/15 · website intake 12/12 | Browser workflow · 10 Oct 2026 |
| Security review | Every finding fixed, each with a regression test | Three independent reviewers |
| Staging | Verified on the real hosting, database and sign-in services | Acceptance · 7 Oct 2026 |
| Production | First release deployed · not yet launched to clients | Sign-in only · launch needs my approval |
06 — My part
I decided what it should do, how it should look and what agents may never do. Claude Code wrote most of the code — and the tests — inside those rules.
- Product scope and the 58-model data design
- The Fable interface language: tokens, type, layout
- Autonomy levels, approval and budget rules
- Release gates and staging acceptance