# Changelog — Sumloop

What shipped and when. Newest first.

## 2026-08-18 — The 30 day trial no longer asks for a card

_Billing_

Start a trial, connect a vendor and see your own numbers without entering payment details at all. Before day 30 Stripe emails you a link to add a card; if you never do, the subscription ends and the workspace locks. Nothing is charged without a card you entered yourself.

This reverses a deliberate decision, so the reasoning is worth stating. A card up front makes day 31 a renewal rather than a sale to win again, which is a real advantage and is now given up. What outweighed it is who never reaches day 1: a card is what someone enters once they have decided, and a trial exists for someone who has not. Asking first does not slow a sceptic down, it removes them — and it flatters trial-to-paid conversion by admitting only the already-sold, so the number that would have revealed the problem is the one the policy improved.

There is also a new mix sumloop.stripe, which compares the sandbox and production Stripe setups against each other, against the prices in this repo, and against the permissions the app's key actually needs. The two are separate accounts — a price, webhook endpoint or portal configuration created in one does not exist in the other — and nothing in Stripe compares them for you. Its first run found production with no customer portal configuration at all, which would have meant every “Manage billing” click failing and nobody being able to cancel. Fixed the same day, before anyone met it.

## 2026-08-18 — About and security pages, and a landing page reordered for the sceptic

_Platform_

There are now /about and /security pages. Nothing on either is new — where the data lives, what a connected credential can and cannot do, who else processes it, what we do not claim — it was all in the privacy policy or the integration guides, which is not where someone deciding whether to trust a young product goes looking.

The landing page changed order. The section that says “you are about to give a young product your billing credentials” used to be eight screens down, below the vendor list and the amortization explainer; it is now the third thing you read. It also leads with the answer: three of the four ways a cost gets in touch no vendor account at all, so a workspace can run without handing us a single credential.

“No spreadsheet to maintain” is gone. It told people who had built one that their work was bad, which is a poor way to win an argument. The line is now “a spreadsheet is only true the day you update it” — the same point, aimed at the failure mode instead of the person.

The demo now catches something. Its cloud bill jumps a third in June and comes back in July, and both movements are in the variance table rather than only on the chart — because the only question a visitor is really asking is whether this would have found anything.

## 2026-08-18 — The demo now runs the forecast and the export, not just the overview

_Product_

/demo has three sections rather than one: the overview, the spend forecast, and the export builder. Each is the panel a paying workspace uses, not a screenshot of it — the same components, served an invented workspace over a public read-only endpoint.

The forecast runs the real models. Every service is backtested against seasonal naive, moving average, linear trend, Holt and Holt-Winters, and whichever tracked that service's own history best is the one kept — which is why a seat-priced bill that stepped when the team grew is not fitted the same way as a cloud bill that climbs.

The export is a real export. Five datasets, filters validated by the server, a row count that is the count of the file, and a CSV or JSON you can actually download. No signup, no card.

One fix fell out of building it: with exactly twelve months of history, seasonal naive was scored during backtesting as “repeat the last value” and then forecast as “replay the whole year” — one model measured, a different one shipped. It now waits for more than one full season before claiming a season, so what wins the backtest is what you get.

## 2026-08-18 — Twenty-seven new integrations, and multiple accounts per provider

_Integrations, Platform_

Sumloop now syncs 49 vendors. The new ones: GitHub, Google Cloud, Vercel, Microsoft 365, Slack, Notion, Linear, Sentry, Auth0, Clerk, Algolia, Neon, Upstash, bunny.net, Cloudinary, Mailgun, Zoom, Tailscale, Bitwarden, Dropbox Business, Intercom, incident.io, Ably, PagerDuty, LaunchDarkly, Postmark and Ahrefs.

A workspace can now connect the same provider more than once — two AWS accounts, a second GitHub org, a staging Stripe. Each connection is named and carries its own billing cycle.

Google Cloud was the hardest of these: Google publishes no cost endpoint at all, so Sumloop reads the BigQuery billing export from a service account and sums cost plus credits the way Google itself defines net cost.

## 2026-08-18 — Every page is now readable by an AI agent

_Platform_

Sumloop's marketing pages were a JavaScript shell: correct for Google, empty for ChatGPT, Claude and Perplexity, whose fetchers do not run JavaScript.

Every public page now also answers as markdown at its own address with .md appended, indexed from /llms.txt. Setup guides carry HowTo structured data and the pricing questions carry FAQPage, so an assistant asked to compare cost tools can read what Sumloop does, what it connects to and what it charges — without rendering anything.