January 2026

Atlas

Nightly Claude analysis across ~270 tickers, a signals engine, and a weekly pick.

Live ↗

marketsagents

What it is

Atlas — stonk.world — is market intelligence for the AI, quantum, and space supply chains, built for retail investors like me rather than for desks. A universe of roughly 270 active names across a dozen sectors gets re-analyzed every night, and the site is where that lands.

I have a normal job. Atlas exists because the research I wanted to read didn’t exist at a price a retail investor would pay.

What actually runs

A stack of scheduled jobs, each with a narrow responsibility:

• Nightly analysis — every ticker in the main universe gets a fresh Claude pass after the close: what changed, whether the thesis still holds, whether conviction moved.
• The signals engine — insider Form 4 buys and sells, 8-K filings, unusual options activity, datacenter power deals, executive transitions. A scoring pass then looks for corroboration: a single insider buy is noise, but an insider buy plus an 8-K plus options flow on the same name in the same week is a constellation worth surfacing.
• Pick of the Week — a verification-grade weekly pick with hard rejection gates (bankruptcy risk, closed catalyst, pure short squeeze), tracked honestly through the week and closed out on Friday whether it worked or not.
• A weekly briefing — stories distributed across inflection, capital flow, earnings tells, and quiet signals, with a hard quota that at least two lead with a small-cap.

Things I got wrong first

Seed data is not a fallback. The ticker tape carried 28 hardcoded change percentages it silently fell back on whenever the price API missed a symbol. A dead feed and a live one looked identical — for months. The rule now: if a live source exists and returns nothing, render an em-dash. An invented number hides the failure and removes all pressure to fix it.

A broken pipeline is not an excuse to build a second one. When the options data source died, the fix was to migrate it, not to quietly stand up a parallel endpoint next to the broken one. Two pipes where one belongs is worse than an obvious outage, because nothing looks wrong.

Own your own data. Daily bars for the whole universe now live in my own store — about a decade deep, split-adjusted, appended nightly — rather than being re-fetched from a vendor on every request.

Stack

React and Vite on Vercel, Supabase Postgres, Clerk for auth, Polygon for market data, and the Anthropic API doing the analysis. Plus a free undocumented Nasdaq endpoint that turns out to serve ten years of split-adjusted history, full option chains, short interest, and institutional holdings — with no key.