Give an AI assistant the ability to publish your backlog

MCP, the Model Context Protocol, is the standard way to give an AI assistant a tool it can actually use rather than describe. Instead of an assistant explaining how you might convert your posts, it converts them, because the conversion is a tool it can call and the result is a file rather than a suggestion.

The engine underneath is the same one the browser applet uses. That was a deliberate design decision made early and it is the reason this page exists: the conversion is a pure function with no I/O and no environment specific behaviour, so the browser, the command line and the MCP server are three thin wrappers around one piece of logic. Fix a parsing bug once and all three are fixed.

That structure is worth more than the applet it produced. Most tools of this kind grow a second implementation the moment they need a second interface, and from then on the two drift, and the bugs get fixed in whichever one somebody was looking at. Keeping the engine pure is what prevents that, and it costs nothing at the point you decide it.

Self-hosting is the other deliberate choice. This is a server you run, not a service you send your content to, which for an unpublished backlog is the difference between a tool and a disclosure. The setup is written out on this page rather than assumed.

It is worth being clear about who this is for. If you are converting one blog once, use the applet and close the tab. This is for the case where publishing is a repeated pipeline, where an assistant is already part of how the work gets done, and where the conversion should happen inside that flow rather than as a manual detour through a browser.

As a piece of Labs work it says something specific: that the interesting decision was made in the architecture rather than in the interface, and that a tool built as a pure function with wrappers around it can be given a new front end years later without touching the part that does the work.