MCP markdown export: preserve block reference provenance and date entity structure

I'm a daily Capacities user running a structured knowledge system (MOCs, inquiry notes, a hybrid PARA and Zettelkasten-style slipbox system) that feeds a heavy writing workflow: position papers, board material, talks, and long-form strategic work. For the AI side of that workflow I use Claude, connected to my space through the Capacities MCP integration. Claude has become my primary thinking and writing partner over this system precisely because the MCP connection lets it read my actual objects rather than pasted fragments, and for me that difference is everything.

Today I ran deliberate experiments with Claude on what survives the markdown export, testing from both directions with live objects. Two gaps significantly limit the workflow, and both look like serialisation choices rather than hard problems, since the structure already exists in the app.

1. Block references lose their source pointer entirely.

When Claude fetches an object containing a block reference, the transcluded text arrives as plain text, which is good, but the reference structure is completely flattened. No link, no URL, no object ID pointing to the source. The source object likewise carries no marker that one of its blocks is referenced elsewhere. We verified this empirically in both directions: a connection that is first-class inside the app is invisible across the MCP boundary.

Practical impact: block references are the natural way to mint a permanent note from a journal or reading note, but Claude cannot trace that provenance when helping me draft from my notes. My workaround is hand-writing a "created from:" object link into every promoted note, which works but duplicates structure Capacities already holds.

Request: when rendering a block reference to markdown, emit the transcluded text plus a machine-readable source pointer, for example a trailing link containing the source object ID (and ideally a block ID). Even a simple suffix like [source](https://app.capacities.io/{spaceId}/{objectId}) after the transcluded text would make the knowledge graph traversable by Claude.

2. Date entities flatten to unstructured prose.

A date inserted with the date function exports to Claude as plain human-formatted text, e.g. "July 7, 2026", with no link and no ISO form. The calendar connection, one of Capacities' most distinctive structures, vanishes at the boundary. My workaround is typing ISO dates manually into headings so Claude can reason about the chronology of my journal entries.

Request: export date entities as links with an ISO 8601 date in the URL or title, so AI tools can sort, filter, and reason about time.

Related, lower priority: backlinks and the Related Content panel are also invisible over MCP. Exposing backlinks, either in the export or via a dedicated MCP tool, would let Claude see what points at an object, not just what it points to.

Capacities' object model is genuinely well suited to AI-assisted writing and sets an amazing platform for RAG; these fixes would let the export carry the structure the app already has. As it stands, users who want AI-traversable provenance must hand-code it. I hope that future development of the MCP connection can take this into consideration.

I am finding the Capacities / Readwise / Claude combination to be potent and powerful for my use case.