A site has a life beyond launch. sitekit is built around four loops, each following the same three roles — the agent authors, the CLI aligns, providers act. Your agent reaches for the right one when you describe the task in plain language; you don’t have to name the commands.
Once a page is authored, never re-run sitekit run on it — that overwrites the body with a blank scaffold. Editing an existing site is the maintain loop, not a rebuild.

Build

Create a site and ship it.
sitekit init <name> --domain <d>.app --description "<pitch>"
sitekit run --shape=marketing-site --style=editorial   # the agent authors each page
sitekit verify                                          # iterate to clean
sitekit deploy --env staging                            # then --env prod
See the Quickstart for the full first run.

Maintain

Edit an existing site surgically — change a section, fix a broken link, restyle, or see what’s drifted — without re-running the agent on a whole page.
sitekit inspect                  # read the site model + detect drift (no agent)
sitekit reconcile                # apply a surgical edit
sitekit verify --maintenance     # the maintenance-specific checks
Page lifecycle lives here too: sitekit page rename | move | delete renames, moves, or deletes a page while rewriting inbound links and the content binding. reconcile --restamp re-stamps pages after a pack update with no agent call, and the visual pass (verify --visual) flags before/after regressions once you’ve approved a baseline.

Experiment

Run an on-brand A/B test and ship the winner you choose.
sitekit experiment new --section <id>
sitekit experiment variant --section <id> --from <fragment.html>
sitekit experiment start          # fail-closed conformance gate + the traffic splitter
sitekit experiment status         # honest two-proportion stats — facts, with sample size
sitekit experiment promote        # you decide; sitekit applies it safely
sitekit measures and applies; it never declares a winner for you. It reports the numbers with the sample size attached and leaves the call to you.

Evolve

Feed a business change — a new feature, a sharper pitch, a reposition — forward into the live site, rewriting only what’s affected.
sitekit brand evolve              # snapshot the brand source + install the evolve skill
# the agent mutates the voice pack + content/messaging.yaml
sitekit brand record              # validate + bump the voice version + report stale pages
sitekit reconcile --brand-delta   # cascade the change across affected pages

More capabilities

Add a page

Add new real estate after the first build:
sitekit site add-page <slug> --shape=<name>
sitekit records the page in the site’s compound contract and emits its scaffold. The agent authors the body and wires the nav link; sitekit verify flags the page as unreachable until it’s linked.

Manage copy

Edit just the words — no markup — through a plain, copywriter-friendly view:
sitekit content export            # write content/copy/<page>.md
# edit the Markdown
sitekit content import            # deterministically swap the words back into the HTML
sitekit verify
Plain-text fields swap deterministically; the agent re-authors anything that’s markup. The same copy layer backs the front-end editor (sitekit content set / apply, and the localhost sitekit edit-bridge).

Adopt and maintain a foreign site

sitekit works on sites it never built — WordPress, Squarespace, a custom stack. Capture the brand, then keep the site on-brand without rebuilding or moving it:
sitekit adopt <name> --voice=<pack> --style=<pack>   # record the brand binding
sitekit align --url https://example.com/page         # judge a live page against the brand
sitekit align --fragment ./edit.html                 # judge an edit before it ships
sitekit is the conscience — it records, verifies, and reports. The agent is the hands — it authors the change and writes it back into the foreign stack. The same portable engine can run conversion experiments on a foreign site through a single embed.

Serve the site to AI agents

Project the site into the channels other agents read, on-brand by construction:
sitekit agents emit               # write pages/llms.txt + inject a Schema.org block
sitekit agents check              # the agent-discovery checks
Both are pure projections from your content — no agent call. Re-emit after content changes.

How it works

The three roles every loop shares.

Command reference

Every command group, by task.