RFC 5023 (2007), the editing half of Atom — collections you POST entries into, service documents that describe them, ETag-conditioned updates, and media resources whose binary content is posted separately from the entry that describes it. It is the standards-track origin of the “metadata and bytes are different resources” pattern that every modern upload API uses, and it was the HTTP API that Google, Microsoft and the blogging platforms built on in the late 2000s.
Atom Publishing Protocol
The Atom Publishing Protocol is what you got when the people who had just finished the Atom feed format asked how a client should create and edit the entries in one. The answer was a small, careful use of HTTP: a service document lists collections, POST to a collection creates a member and returns its location, PUT replaces it, DELETE removes it, and ETag with If-Match guards every update. Nothing in it is exotic, which is the point.
- Collections and service documents - Discovery is a document, not documentation. A client reads the service document and knows where it may post and what media types each collection accepts.
- Media resources and media link entries - Posting a binary to a collection creates two things: the media resource holding the bytes and a media link entry holding its metadata. Edit either independently.
- Conditional updates - Servers return ETags and clients send
If-Match; a lost update is a412, not a surprise. - Extensible by category and link - Categories documents and link relations carry the domain-specific parts without changing the protocol.
AtomPub mattered because for several years it was the web API: Google’s GData, Microsoft’s early OData, and most blogging platforms spoke it before JSON won. It is in this catalog as the earliest standards-track statement of a pattern that is still being re-specified. When the Model Context Protocol Files Working Group agreed in October 2026 that an upload should send metadata first, get a target back, then send bytes, they described the media link entry from 2007. CMIS carried an AtomPub binding for the same reason. The protocol is dead in practice; the design survived everywhere.
