install · v1.7.0
Two lines, inside any Claude Code session
Add the marketplace, install the plugin. Everything after that happens in the session you are already sitting in.
/plugin marketplace add blyatiful1/ultraweb/plugin install ultraweb@ultrawebthen, in the same session
/reload-pluginslater, when a new version lands
/plugin marketplace update ultrawebConfirm the dialog, reload, done.
did it work
Three ways to know it took
/plugin lists ultraweb
The marketplace entry is registered and the plugin is enabled — the install is not a promise, it is a row in a list.
/ultraweb answers when called
The orchestrator responds instead of the session improvising a website from vibes.
Phase 0 re-verifies before it spends
The build's own preflight checks the install, the toolchain, and the browser again before any expensive work starts — a dead end costs two minutes, not five hours.
what you need
Four requirements and one nicety
| Claude Code | required | CLI, desktop, or web.required |
|---|---|---|
| Node + npm | required | Something has to build the Next.js app.required |
| Playwright MCP | required | The eyes. Without it the visual, responsive, and accessibility gates record an honest UNVERIFIED instead of quietly waving the site through — declared in Phase 0, never discovered five hours in.required |
| Vercel CLI auth | optional | Preview URLs at the review checkpoints. Without it you review screenshots, and nothing blocks.optional |
| token headroom | required | A full run is a studio engagement, not an API call.required |
the honest bill
Priced in your minutes, then in tokens
Two dials decide what a run costs you, and they charge in different currencies.
The engagement dial costs minutes of your time
| dial | blocking checkpoints | your time |
|---|---|---|
| hands-offnobody has to be at the desk | none | 0 min |
| guidedthe default | 3 | ≈ 30 min |
| studiosign-off at every milestone | 6 | ≈ 43 min |
The guided thirty, itemised: direction approval ~10 min, homepage review ~5, acceptance ~15. Every checkpoint announces its own number before it takes the time, and an unattended one auto-passes rather than stalling the build.
The scope dial sizes the pipeline
- sketch
- Two mockup candidates, a thinner pipeline, a fraction of a full run.
- standard
- The default: the full twelve phases.
- flagship
- Adds the showpiece budget and up to five visual-critique rounds.
Read that as sizing, not pricing — it is the README's own back-of-envelope estimate. The plugin refuses to invent precise numbers for a site it has not built yet, and neither does this page.
How to spend less
Say what the job really is: “landing page” picks the sketch tier before the first token is spent. Then build once and talk to the built site — everything after the first run is iterate, which touches only what your change touched.
One earlier build of this site's predecessor was measured end to end — 2026-07-28, classic autonomous mode, back when the harness had 72 skills. Its full receipt, counted from session transcripts, lives in the plugin README.
the measured receiptcheaper first tastes
Two ways in that are not a build
retrofit
the zero-risk first taste
Point it at any Next.js site and read back a scored, unflattering gap report — every gap naming the skill that would fix it. No build, no big bill.
iterate
the surgical pass
For a site ultraweb built: say what is wrong — the hero is too timid — and it scopes the change, then re-runs only the gates you disturbed.
Working offline, or from a fork
Clone it anywhere and add the clone as a local marketplace — same namespacing, same 80 skills.
git clone https://github.com/blyatiful1/ultraweb.git ~/src/ultraweb/plugin marketplace add ~/src/ultraweb/plugin install ultraweb@ultrawebDo not clone straight into ~/.claude/skills/ — a nested skills tree half-installs, and preflight will say so.