The Best Tools Show Up Before You Go Looking
One of Atlas's background jobs is a weekly scan of GitHub's trending repositories. It is not there to keep me entertained. It is looking for tools and libraries that could plausibly fit into Atlas's own codebase, and when it finds one, it does not act on it. It writes the candidate to a suggestions page and waits for me to browse.
Yesterday I was scrolling that page and stopped on a repo called Archify.
The tool I did not know I needed
Archify is a Claude Skill that generates architecture diagrams: system diagrams, sequence diagrams, data flows, lifecycle charts. You describe your system in plain language, and it hands back a single self-contained HTML file with a dark and light theme toggle, 4x resolution export, and copy to clipboard built in.
No Mermaid syntax to fight. No draw.io. No dragging boxes around a canvas for 45 minutes and then nudging arrows because one of them looks slightly off.
I did not go looking for this. I did not know it existed, and honestly I had made a kind of peace with diagramming being tedious. Atlas found it, decided it was relevant to the way I work, and put it in front of me. The tool showed up before I went looking, which is the only reason I found it at all.
The part of local AI nobody talks about
When people picture an AI assistant, they picture a prompt and a reply. You ask, it answers. That framing quietly assumes you already know what to ask, which means you only ever discover what you already knew to look for.
The more useful thing an assistant can do is run while you are not asking. Atlas scans GitHub trending every week, evaluates repos against what it already knows about its own architecture, proposes the ones that fit, and lets me approve or dismiss each one. Archify came through that pipeline. So do the ideas I glance at and reject, which is most of them, and rejecting them is the point too. The suggestions page is only valuable because I trust that nothing on it got there to manipulate me.
That is the difference I keep coming back to. A recommendation feed on someone else's platform is optimizing for engagement: what keeps you scrolling, what sells, what the algorithm was told to push this quarter. Atlas's suggestions page is optimizing for one thing, whether a tool is genuinely useful to me, because there is no one else in the loop to optimize for anything else.
No subscription. No engagement metric. No feed tuned by a company whose incentives are not mine. Just a local model on a $1,000 box doing quiet recon on my behalf, every week, and reporting back.
Surfacing is the feature
I have written before about Atlas running millions of tokens a month on small background jobs. This is what a lot of those jobs are actually for. Not answering questions I pose, but noticing things I would never have thought to ask about, and surfacing the few that matter.
Archify is a small, concrete example of a larger pattern I did not fully appreciate until I had built it: the best tools tend to show up before you go looking for them, but only if something is looking on your behalf. Search finds what you already know to search for. An agent doing recon in the background finds the thing you did not know was missing.
If you want to see the tool that kicked this off, Archify is open source at github.com/tt-a1i/archify.
The question I will leave you with is the one this made me ask myself: what is the most useful thing an AI has surfaced for you that you were not searching for?