The Tools Behind AutoShort: Our Real Stack and What Each One Costs

Some links on this page are paid or referral links (Claude), marked next to each link. If you sign up through one, we may earn a commission or usage credit at no extra cost to you. We only add paid links for tools we use ourselves.

Why we publish this

“What tools do you use?” is the question we get most from other founders and from creators who want to understand how AutoShort works. Most stack lists online are affiliate catalogues: forty tools, each with a link, none obviously used. This one is short because our stack is short. Every tool below is something we use this month, with what we pay where we know the number exactly. Where we have an affiliate or referral link, it is marked next to the link.

The rule we follow, and recommend: add a tool only when its absence is costing you time every week. Most of the tools we tried and dropped solved problems we did not have yet.

Building: Claude Code

AutoShort and its sister product StoryShort are built by a very small team, and Claude (referral link: 7-day guest pass; we may receive usage credit) Code does most of the typing. We are on the Max plan at $200 a month, which is our single biggest software cost and, by a wide margin, the best value. It writes features, writes the tests that check them, reviews pull requests and runs the boring operational chores from the terminal.

Two habits make it work. First, a CLAUDE.md in each repo that records the rules we learned the hard way, so the next session does not relearn them. Second, nothing merges on Claude's word alone: tests run on the merged code, and a second reviewer (another agent or a human) reads the diff. Claude is quick and confident, and confidence is not the same as correct.

It also drives the product itself: AutoShort has an MCP server, so Claude can plan and render videos. The full walkthrough shows how.

Hosting: one droplet

Our products, the shared video worker and the databases run on a single DigitalOcean droplet with 16 GB of RAM and 320 GB of disk. We deploy with Kamal, which builds a Docker image and swaps containers with no downtime. There is no Kubernetes, no serverless bill and no managed database.

It started as a smaller droplet and has been resized since; a bigger single machine was simpler than splitting into services. The trade-off is real: one machine is one point of failure, and a memory-hungry job can slow everything else down. We accept that at our size because it keeps costs predictable and debugging simple. The day a single box is the bottleneck, we will split the worker out first, since rendering is the heavy part.

Voices: Google and ElevenLabs

Every AutoShort video is narrated. The default is Google Cloud text-to-speech, billed per character with a monthly free allowance on its standard natural voices. Creators who want a premium voice can choose ElevenLabs, and we pay for its Creator plan at $22 a month to serve them.

We used to assume the premium voice was what kept customers happy. Our own ratings data said otherwise: videos with Google voices were rated helpful more often. The details and the caveats are in the voice comparison.

Research: vidIQ

For deciding what our own channels should make, we use vidIQ's free tier through its MCP connector, so Claude can run keyword research and look for outlier videos in a niche during the same session where we plan content. The free tier gives a monthly allowance of credits, which has been enough for our volume. We wrote up what it helped with and what it did not.

Distribution: Typefully

We build in public on X. Drafting threads in Typefully and scheduling them for the hours our audience is awake beats posting live at midnight our time. We are on the free plan, which caps published posts per month; that cap is a useful discipline. Our process is in the Typefully post.

Three lessons from keeping it small

Free tiers are a feature, not a phase. vidIQ and Typefully both cap the free plan, and both caps turned out to match how much we should be doing anyway. When a cap starts to hurt every single month, that is the signal to pay, not before.

Measure the bill you think you know. We assumed our Google voice bill was small because the standard voices have a generous free allowance. Then a newer, pricier voice became the default in one code path and the bill quietly grew. A monthly look at every invoice, line by line, is the cheapest audit there is.

Prefer tools an agent can drive. Most of what we added this year has an API or an MCP connector, so Claude can use it inside the same session where we plan or build. A tool that only works by clicking through a dashboard costs a person's attention every time.

The whole bill

Missing from the table on purpose: the per-video costs of image generation and video models. They scale with how many videos customers make, and they are part of what a credit pays for, not a fixed tool subscription.

What we do not use

We do not recommend other AI video generators here, for an obvious reason: we make one. If you are comparing, our comparison pages lay out the differences with sources. We also do not use paid SEO suites, a CRM or a no-code automation platform yet. Each of those is a fine product; none of them is solving a weekly problem for us today.

If you are starting a faceless channel rather than a software company, your stack is even smaller: one tool to make the videos, one to research ideas, and your own judgment on the hook. Everything else can wait until it hurts.