16 capabilities, grouped by where they sit in the pipeline

Cost does not scale with page count. One page or a hundred thousand.

A model migrates one to three exemplar pages interactively. A human approves. The agent then writes a deterministic TypeScript importer module with fixture tests — and that module, not the model, runs over the remaining pages: in parallel, resumable, at zero model cost.

A defect found on page 4,312 is fixed in the importer, so it is fixed for all 10,000 — and the fix is a failing unit test, not a re-roll of a stochastic process.

Enforced by the dependency graph, not by a policy someone has to remember.

The six packages that do the bulk work — UCM, inference, importer, crawler, emitters, verification — list no dependency by which a model call could be reached. Their entire test suite runs with no network and no credential. A violation would be a package.json change, visible in review.

Pixel-match the old site, or map it into the library you already have.

PARITY reproduces the source. DESIGN_SYSTEM maps every region into an existing Figma-derived library and refuses to invent a block that is not in it. HYBRID lets the two coexist per region, which is what a real migration usually needs.

The same design on two pages becomes one component, not two.

A homepage hero and an about-page banner carrying the same fields are one component with different content — that is what a component is for. A duplicate is refused outright, with the existing name to use instead. The difference is a site where a change is made once and a site where it is made in fourteen places and missed in three.

The verification report shows how many pages use each component. One used on a single page is usually two designs that should have been one.

One block definition emits Edge Delivery and the Universal Editor.

A single BlockType produces the EDS block, its CSS and JS, and the AEM Cloud Service / Universal Editor model, authoring table and component-definition.json. AEM Cloud is not an afterthought bolted on later — the two targets are generated from the same source of truth.

The blocks your migrated content needs to render, generated with it.

Most migration tools move content into components someone else has to build. UEMA generates the blocks, the CSS, the JS and the Universal Editor models — and validates them against the platform contract: no build step, no runtime dependencies, scoped CSS.

The contract is enforced by a validator, not requested in a prompt.

What you review is the commit, not a preview of an intention.

Emission assembles the artifacts and diffs them against your branch. Stage what you want, and UEMA commits, pushes and opens — or reuses — a pull request through UEMA Connect, a dedicated GitHub App rather than a personal access token. A deleted path stays visible with its prior content until the next emission, so the deletion is reviewable too.

Every commit is attributed to the App and its tokens expire in an hour — more on that below.

A migrated site that still serves its images from the old one is not migrated.

Pictures are copied across, converted to a modern format, and stored once however many pages use them. Images served from another company's domain are listed rather than copied, because re-hosting someone else's media is a licensing question nobody asked.

Deduplicated by content, so a logo on every page of a 5,000-page site is fetched once and stored once.

Carried across as site-wide pieces, not repeated on every page.

The header, menu and footer come over as shared pieces your authors can edit in one place — which is how they are meant to work, and what keeps them editable after the migration is finished.

Where pages disagree, the version most of them carry wins, and the count is reported so genuinely sectioned navigation is visible rather than flattened.

Thirteen AI providers, one setting — and none of them in the bulk path.

Claude, Gemini, OpenAI, Azure OpenAI, Groq, Mistral, Cerebras, OpenRouter, Together AI, GitHub Models, Hugging Face, Cohere and Cloudflare Workers AI — plus any OpenAI-compatible endpoint, including one running on your own hardware. Use whichever account you already have, switch whenever you like, and route different steps to different providers if a cheap model is good enough for one of them.

Switching provider is a dropdown, not a migration. Adding a credential makes a provider available and reassigns nothing on its own.

Your migration is yours, and the boundary is tested.

Tenants derive from your email domain. Three roles, a rule that no path can violate — a tenant's last active owner cannot be suspended or removed — and every route naming a project resolves it within your tenant and answers 404 otherwise. An unverified address at a company domain confers no membership at all.

A change to labelling is scored, not argued about.

The places a model is used — region labelling, block mapping, importer authoring, style import — are scored against a fixed corpus, so a prompt or model change has a number attached rather than a vibe. Adapters exist for the importer, the block mapper and the style importer.