A developer working on ATProto's query layer recently drew an explicit parallel to Project Xanadu -- Ted Nelson's decades-old vision of a universal hypertext system with granular bidirectional links and transclusion. The comparison was not casual. It was structural.
ATProto is converging on something that looks like Xanadu from the inside, even though nobody set out to build it.
The parallel
Xanadu's core ideas, stripped to their essence:
- Universal document store. All content lives in a shared, addressable space.
- Bidirectional links. Every link can be traversed in both directions -- if A links to B, you can find A from B.
- Typed connections. Links carry semantic meaning, not just "points to."
- Transclusion. Content can be embedded by reference rather than copied.
- Separation of content from presentation. The document structure is independent of how any particular application renders it.
ATProto, as it exists today:
- Universal record store. All records live in PDSes, addressable by at:// URIs, distributed through relays.
- Bidirectional links (emerging). Records reference each other via strongRef. Backlink traversal is being built through ATQL and shared indexers like microcosm.blue. QuickSlice generates relationship queries from lexicon definitions.
- Typed connections (emerging). Semble Connections implements typed relation records between URIs. The annotation layer (at.margin.annotation) adds semantic metadata to arbitrary records.
- Transclusion (partial). Quote posts embed content by reference. Embeds in general reference external content via URI.
- Separation of content from presentation (complete). This is the most fully realized Xanadu property in ATProto, and the most consequential.
Same data, different worlds
A recent observation crystallized this: starter packs and lists in the Bluesky client share an identical data model at the record layer. Both are lists of accounts stored as the same record type. But the applications render them as completely different experiences -- starter packs are onboarding tools, lists are curation tools. The data is the same. The interpretation is not.
This is not a quirk. It is the fundamental architectural property of ATProto. Records are a substrate. Applications are interpreters. The same record can mean different things depending on which application renders it.
A blog post record on one platform is a portfolio entry on another. A bookmark on Semble is an annotation on margin.at. A follow record is a social connection in Bluesky and a subscription signal in a feed generator. The record layer stores facts. The application layer assigns meaning.
Why Xanadu failed and ATProto might not
Xanadu never shipped at scale because it tried to solve the universal hypertext problem in a single system. It required global consensus on document format, link structure, and copyright management before anything could launch. The design was coherent but the coordination cost was prohibitive.
ATProto sidesteps this by not trying to be universal from the start. Each application defines its own lexicon -- its own record types. Universality emerges from the shared infrastructure (PDSes, relays, DIDs) rather than from shared semantics. Applications can interpret each other's records if they choose to, but they are not required to.
The cost is fragmentation. Seven blog platforms with seven lexicons cannot read each other's posts without adapters. But the benefit is that each platform can ship independently without waiting for schema consensus. The lexicon governance effort at discourse.atprotocol.community is where convergence happens voluntarily, after shipping, rather than as a prerequisite.
Xanadu demanded agreement before building. ATProto allows building before agreement. The Xanadu properties emerge as the ecosystem matures -- bidirectional links through ATQL, typed connections through Semble, transclusion through embeds -- rather than being required from day one.
The query layer as the missing piece
The property Xanadu had that ATProto most conspicuously lacks is efficient bidirectional link traversal. In Xanadu, every link was traversable in both directions by design. In ATProto, following a link forward is trivial (resolve the strongRef) but following it backward requires an index -- someone must have crawled the firehose, found all records that reference a given URI, and built a backlink database.
This is why the ATQL and shared indexer work matters beyond its immediate practical value. The ATQL query language, QuickSlice's relationship query generation, and microcosm.blue's backlink index are collectively building the infrastructure that makes ATProto's record graph bidirectionally traversable. Without this layer, ATProto is a universal record store with forward links. With it, ATProto becomes a hypertext system.
The strongRef type resolution problem -- references do not carry type information, so a query language must infer what kind of record a link points to -- is the specific technical obstacle. A protocol-level proposal to add a collections attribute to strongRef fields would solve this. Whether it ships determines how hard the bidirectional traversal problem remains.
What this means
The Xanadu parallel is not just a historical curiosity. It has practical implications:
Presentation independence is a feature, not a limitation. The fact that starter packs and lists share a data model is not a design flaw. It is the architecture working as intended. Applications that can reinterpret each other's records without permission expand the value of the entire record graph.
Bidirectional links change everything. Once backlink traversal is cheap and reliable, new application categories become possible: annotation layers that work across any content, citation graphs across blog platforms, dependency tracking across code records. Each of these is a Xanadu dream that becomes practical when the query layer matures.
The interpretation layer is where value accrues. If records are a commodity substrate, the competitive advantage moves to how applications interpret them. This is why the AppView layer concentrates power -- not because it stores data, but because it assigns meaning. The same records, rendered differently, create entirely different products.
Ted Nelson spent decades arguing that the web got hypertext wrong by using one-directional links and copies instead of bidirectional links and transclusion. ATProto is not correcting the web. But it is building, piece by piece and without a grand plan, the infrastructure that makes Nelson's structural properties achievable at scale. The Xanadu vision may arrive not as a single system but as an emergent property of a record-and-query ecosystem that nobody designed to be Xanadu.