Case study

Keel

An operating system for a one-person business, run by AI agents. They write the code, run the nightly research, score new ideas against a rubric, and log every action. The morning brief tells you the truth about where you stand, including when the answer is unflattering.

Type: Product / agent platform
Role: Design, build, run
Year: 2026
Stack: Next.js · Supabase · Claude · pgvector

One graph, live status

MAP5 ENTITIES LIVE

ListingMate

Marsh Property

Content agent

Research agent

Social automation

productclientagentcapability

Watch it work, 22 seconds

It tells you where you stand

THE MORNING POSITIONSCORED 12H AGO

▌

PIPELINE OPEN

$129/mo

MONTHS TO TARGET

GROWTH

flat, no ETA

Not a dashboard of green numbers. A written position, in plain English, that will tell you the pipeline is worth nothing on paper if that is what the data says.

The problem

Running several products and clients solo means the work scatters across repos, notes, and a dozen browser tabs. Nothing knows about anything else. The important next move is always buried under the noisy one.

What I built

One system over one database, with a clear spine: capabilities become products, products serve clients, projects build them. AI agents run against that spine. They write code on demand, refine ideas every night, and surface the single next move. Every action lands as a reviewable draft.

Idea Refinery, scored every night

IDEA REFINERY0/40

Sell the outreach probe as a standalone lead-list product

Revenue proximity

Build cost

Fit to capability

Evidence of demand

A nightly agent scores each new idea against the rubric, then offers to promote the winners into real projects, products or agents. You wake up to decisions, not a backlog.

Nothing ships without you

MISSION

Decide whether the Cairns trades outreach earns a third batch

3 JOBS

researcherQUEUED

Pull reply and bounce rates from both sent batches

analystQUEUED

Compare against the $8,000/mo target at $299 per sale

writerQUEUED

Draft the recommendation either way

RETURNED, AWAITING YOU

Nothing has run yet.

IT RUNS WHILE YOU SLEEP. IT DOES NOT DECIDE.

Hand over a large task and it is split across a team of sub-agents: planner, researcher, writer, builder. Each returns a draft, and each draft waits on a human decision before anything moves.

The command centre

YOUR MOVE0 WAITING

Nothing yet. The workers are still running.

Every build, brief, sync and nightly run is logged, newest first, with its status. Nothing an agent does is invisible.

Full audit trail

ACTIVITY0 LOGGED

Waiting for the overnight run.

1 screen

the whole business, live and colour-coded by status

Nightly

an agent scores new ideas and proposes the winners

Every run

logged with its output, so no agent action is a black box

The system

Next.js over Supabase with pgvector. A typed data layer both the app and the worker share, so contracts cannot drift. A builder agent runs in an isolated git worktree to write code safely; a nightly worker embeds and scores ideas. Coordination is atomic, with a watchdog that reaps stuck jobs.

The outcome

A scattered business became one legible system that mostly runs itself. Keel is built white-label. The product name, mark and colour come from a single brand file, so the same spine, supervised agents and audit trail can be installed inside another business under its own name.

Want an operating system like this inside your business?

Tell me the job you want to stop doing