An archived project keeps every star it ever earned, so Agentwire reads four flags before it calls anything shipped

An author can archive a project, mark it read only and walk away, and it keeps every star anyone ever gave it. Nobody takes a star back, so the number on that page is a record of attention that already happened, and it says nothing about whether the thing runs today.

The same holds for a template, which is a scaffold published so other people can copy it and fill it in. It holds for a fork of somebody else's work. It holds for a project whose own description says it has not been released yet. All four collect stars exactly the way a finished tool collects them, and a reader deciding what to install has no way to tell them apart from the outside.

Agentwire is our index of other people's shipped agent tooling: MCP servers, coding-agent harnesses and agent frameworks, each credited to whoever built it. The word carrying the most weight in that sentence is shipped, and it is a check in the intake before it is a claim in any sentence on the site.

Watch the Agentwire, MCP servers, coding-agent harnesses and agent frameworks, all of them somebody else's demo

Four flags, read off the project's own page

A candidate has to clear two independent affirmations first. A maintainer added it to a list they publish under their own name, a Hacker News thread above forty points argued about it, its own metadata carries two or more agentic topics, or it passed the star floor for its age. Only then does the shipped check run.

Three of the four things it reads are yes or no fields GitHub already publishes about a project: archived, fork, template. The fourth reads the project's own words, so a description announcing that it is unreleased removes it.

The line we deliberately did not cross was widening that fourth one to catch beta and alpha. A public beta is shipped. Matching those words would have dropped a large part of this whole category for describing itself honestly.

On that run, 137 candidates came through the two-signal test and the shipped check took three of them out. That is a small share, and it is the share we want. The check is aimed at a specific kind of thing, and a version of it that pruned heavily would be catching something else by accident.

What it costs when it does not run

A reader clicks a listing that looks better vouched for than most of what is around it, lands on the page, and finds an author who stopped work months ago or a scaffold with nothing in it yet. The stars gave no warning, because attention accumulates on an intention at the same rate it accumulates on a tool.

These are yes or no fields, so nothing errors if one stops being applied. The index would simply start carrying the things our own pages say it cannot. So the poll prints how many the shipped check dropped on every single run, and a run that reports zero drops with no line at all is one we go and read.

Every entry in the index is running code somebody can clone today: Agentwire.

---

One shipped product, taken apart, once a month. What it does, what it cost to build, what the pipeline behind it looks like, and what the numbers did, read off the repository and the live site, not written from memory. Join the list.