OS OpenShorts GitHub

Decision

Self-host the repo or pay for the cloud: the break-even

The MIT build costs nothing in licence and real money in API calls, time and hardware. The hosted plan is $12/mo for 100 minutes. Here is where each one stops making sense.

The short answer

If you clip a few videos a month and you already own a decent machine, self-hosting is the cheaper and more private option. If you clip every day, or you want clips made without watching them, the hosted cloud is the one that actually saves you time.

What each side gives you

Self-hosted (GitHub repo) Cloud (openshorts.app)
Price $0 for the software, plus your own API calls Free plan, then from $12/mo for 100 minutes
Watermark No Free plan only
Upload cap None Plan minutes
Speed 5–8 min per 8-min video on CPU; ~50 s on an NVIDIA GPU GPU speed, no setup
Your keys Required (Gemini and optionally more) Included
Autopilot (clip every new upload) Not part of the self-hosted stack Yes
API and MCP server Yes Yes
Data Stays on your machine, except what you send to model providers On their infrastructure
You maintain Updates, uptime, disk, security, backups Nothing

The free cloud plan is worth reading carefully before you compare: it clips your first video whole, up to 60 minutes, and after that 20 minutes a month with a watermark. The $12/mo tier is 100 minutes, which is roughly eight 12-minute videos — clip a daily podcast and you will outgrow it in a week. A three-hour stream is nearly two months of that allowance.

The arithmetic nobody does before buying a GPU

Self-hosting is not free, it is metered differently. Three costs sit behind the $0 licence:

  • Model calls. Moment detection scores a transcript, then looks at frames per scene. A one-hour, talk-heavy podcast is the cheap end; a fast-cut video with dozens of scenes is the expensive end.
  • Your time. The stack is Docker plus three containers plus an env file. Getting it up is an evening. Keeping it up is the part people underestimate.
  • Hardware, if you want speed. A CPU run is fine for occasional use. If you clip daily, a GPU at roughly 50 seconds per 8-minute video is the difference between a background job and a bottleneck.

So the honest break-even is: a handful of videos a month on hardware you already own → self-host. Daily volume, or no appetite for maintenance → pay the $12 and stop thinking about it.

Where self-hosting genuinely wins

  • Privacy. Transcripts, sources and finished clips stay on your machine. Point the moment picker at a local model and even the transcript stops leaving the box.
  • Volume without per-minute pricing. Clip a 3-hour stream into 20 shorts and no one counts the minutes.
  • Agent workflows. The repo ships an MCP server and a REST API with keys, so your own automations can drive it — the same surface exists in the cloud, but on your machine you own the queue.
  • No watermark, no plan ceiling. The MIT build has neither.

Where the cloud wins

  • Autopilot. Connect a channel, and every new upload is clipped and scheduled without you touching it. That feature is not in the self-hosted stack.
  • GPU speed with nothing to buy. The same pipeline at roughly 50 seconds per short video, on someone else's GPU, with the model costs inside the subscription instead of three separate dashboards — and no keys to rotate.
  • Nothing to patch. Model releases, dependency updates and the security of a box exposed to the internet are all someone else's problem.

That last point is the whole trade in one line, and it is why this comparison only concerns the project's own two paths. How the self-hosted build compares to OpusClip, Klap and Vizard — which are not open source and have no self-hosted option at any price — is on the alternatives page.

A reasonable middle path

Run the repo locally to evaluate it for a week on short videos, with the free Gemini tier doing the moment picking. That tells you whether the clip quality is good enough for your channel before you either buy a GPU or commit to a subscription. The install guide gets you there in four commands, and the API keys page covers the cheapest configuration that still produces real clips.