The Hub is a private PWA that turns one small server into a studio: chat with an agent, open a real terminal, watch the box — and ship products end to end. Provider-agnostic: any CLI coding agent slots in.
interactive mock · no login, nothing is sent anywhere
02 · The factory
A replay of a real run, with example names — one command provisions everything, the agent builds, and by morning the night shift has already worked a shift of its own.
03 · The loop
The Hub is built as a loop-engineering system: the maker stops being the person who prompts the agent and becomes the person who designs the loop that does. Verification stays human.
every arrow is a gate: a separate judge model reads the stage's report against the real git evidence, and the chain stops at the first stage that doesn't pass. every stage commits its artifact into the repo — the model forgets everything between runs; the repo doesn't.
The pipeline is one pass; the goal loop is the recursion. State a goal once — it's committed as GOAL.md, so the agent re-reads it every cycle — and the Hub goes round on its own. Three models, three jobs: a checker grades the project against the goal and names the one next step, the worker runs it, the judge grades the attempt. Nobody grades their own work.
a loop that cannot stop is a bill, not a feature — it ends on: goal met · cycle cap · judge blocked · an empty cycle · night's end · you pressing stop
IDEA.md → RESEARCH.md → PLAN.md → SPEC.md, committed where the code lives. Any run — tonight's or next month's — picks up exactly where the last one stopped, instead of starting from zero.
The moment work starts it exists as a folder and a dashboard card with its pipeline track — not when it's finished. Open a session in that folder and grind on by hand, any time.
Each stage is checked by a separate, cheaper model that sees only the structured report and the actual diff. The agent that wrote the work never grades its own homework.
Run a stage now while you watch, or queue the chain for the night — with the model picked per task. In the morning the verdict is yours: accept, continue with a different stage or model, or discard.
04 · Everything it does