What Is Still Open

The MCP server works and is published. It is also incomplete in several known ways, and the gaps are deliberate rather than forgotten.

The duplication

The backoffice still runs its own copy of the create logic. It should call the endpoint instead, so there is one implementation rather than two that agree today.

This is deliberately not done. The browser path is what real client work depends on, and it gets switched over once the two are compared side by side.

See Two Implementations for what must stay in agreement meanwhile.

PDF only

create-from-source supports PDF. Markdown and web sources return a 400 with a message pointing at the backoffice:

The ‘{sourceType}’ source type cannot be imported through this endpoint yet. Use the backoffice for markdown and web sources.

One source type was enough to prove the shape.

Note the asymmetry: the capability exists elsewhere in the codebase. WorkflowController fully supports all three, using IHtmlExtractionService for web and IMarkdownExtractionService for markdown. They are simply not wired into the create endpoint.

The request model’s own documentation says "pdf", "markdown", "web", describing the intent rather than the implementation. That is the cheapest item on the list: the same code path with a different extractor.

Two tools of forty-two endpoints

The generated client exposes 42 operations. Two are wired to tools.

That is a position rather than an oversight. A tool is a contract to maintain, and each one should earn its place. But the ratio is worth knowing when asking what the server could do.

Open issues

Issue	What
#131	Dry-run import. Run extraction and return what would land in each field, creating nothing. Lets a session check a PDF before committing, and diagnose a bad mapping without creating and binning a page.
#133	Import report. Which mappings resolved and which did not. `mappedValueCount` is returned already, but a count without its denominator is not much use, since “19” only means something next to “of 19”.
#139	Stale mapping. `extras-to-your-tour.summary` points at a section part `transform.json` does not produce. Surfaced by `list-workflows`. May be the mapping that was meant to fix the price-caveat placement problem, so worth understanding before deleting.
#144	Tool wish list. Running record of what to build next and what deliberately not to.
#152	Stale references. Scaffold residue in `mcp/`, and four frontend comments claiming the MCP server shares their code.

Candidate tools, not committed

From the wish list, with the reasoning that keeps them on it:

validate-workflow reports a workflow’s problems in detail. list-workflows already surfaces warnings, so this is depth rather than new capability. Useful when an import produced less than expected and the question is whether the workflow or the source is at fault.

find-media searches the media library, returning size alongside the key. The specific job is “is this PDF already here, and is it the same revision?” That matters: nine brochures were bulk-seeded ahead of their tours and one was a stale revision, 43 bytes off the live file. A search hit is not proof the right file is there.

Open question whether find-media belongs in UpDoc’s server at all. It is a media-library concern, and Umbraco’s own server may be the right home.

Known defects

Tracked on #152, summarised here because they mislead anyone reading the source:

The server identifies itself as my-umbraco-mcp, the scaffold’s default, in four places
Three files still reference get-chained-info, which was removed, including a live eval test
Four frontend comments claim the MCP server shares create-from-source.ts. It does not.
Four unused MY_* environment variables, template mock code, and two dead entries in the mode registry

What will not be built

From the original design decisions, still holding:

Multi-CMS support. Umbraco-only, forever.
Batch imports. One page at a time with human review between each, deliberately.
Publishing from the tool. The endpoint creates a draft. Publishing stays a separate decision with a human in the loop.

Client-facing use

Running this against a live client site from a desktop AI client is real but deliberately deferred. It needs packaged guardrails, a tightly scoped API user, and a human eye in the loop.

This page describes the state as at September 2026. Check the linked issues for current status.