Shortcast: A Native Mac App for Cutting Video Into Shorts, Fully Local

Most tools that turn a long video into TikTok- or Reels-ready shorts work the same way: upload the footage to a hosted service, wait for a cloud pipeline to process it, download the result. Shortcast, an open-source project from developer mutonby, is a native Mac app for cutting video into shorts that skips the upload step entirely. It transcribes, finds the best moments, cuts clips, writes captions, and reframes to vertical, all running on the Mac itself, with nothing sent anywhere until the operator chooses to publish.

What "native" actually means here

The claim is easy to make and rarely checked. Shortcast's repository backs it up: the app is written in Swift 6 with SwiftUI, not Electron wrapping a web view and not a Python script with a GUI bolted on. Video cutting and export run through AVFoundation, Apple's own media framework. Face tracking for the vertical reframe runs through Vision, Apple's on-device computer vision framework. Model inference runs through MLX, Apple's machine learning framework built for Metal and the Neural Engine. The project's own README puts it plainly: "No Python. No Electron. No embedded runtime. Just Swift, MLX, AVFoundation, Vision."

That stack is also why the app is Apple Silicon-only. It requires an M-series Mac and macOS 15 or later, and it won't run on Intel Macs at all. Building from source needs Xcode 16 and XcodeGen. This is a deliberate, narrow bet on Apple's own hardware and frameworks rather than a cross-platform tool that happens to run on a Mac.

How a long video becomes shorts

The pipeline runs in one pass. WhisperKit's large-v3 model transcribes the source video on-device (or uses an existing .srt/.vtt file if one exists). That transcript goes to a "Director" model, either Gemma 4 12B or Qwen 3.5 9B running locally through MLX, which reads the whole thing and returns the best three to six moments along with a full caption package for TikTok, Instagram, and YouTube in a single response. AVFoundation cuts each moment into its own clip. If the source footage is horizontal, Vision tracks the speaker's face across the clip and an AVMutableVideoComposition pans a vertical 9:16 crop to follow them, falling back to a blurred letterbox when no clear face is found. The results land in a review grid where clips can be played with sound, edited, downloaded, or approved.

There's a second, smaller mode too: drop an already-vertical short and Shortcast will watch and listen to it (using Gemma 4 E4B, a multimodal model whose audio encoder covers the first 30 seconds) and write captions for it directly.

Where the local-first claim has real edges

The privacy pitch holds up under scrutiny, with two honest caveats the project states itself. First, models have to come from somewhere: the Director model (roughly 5 to 7 GB) and WhisperKit download once from Hugging Face on first use, then everything runs offline. Second, publishing is not local. Scheduling or posting a finished short goes through Upload-Post, a third-party API that requires its own account and API key, and the rendered clip is uploaded to that service at the moment of publishing, not before. Shortcast is honest about this boundary in its own documentation rather than blurring it, which is worth crediting. Generating shorts works without ever touching Upload-Post; only the publish and schedule step needs it.

One more real limitation: the app currently ships unsigned and unnotarized, so macOS Gatekeeper blocks it on first open. The README documents the fix, an xattr -dr com.apple.quarantine command run once after installing, but it's a genuine friction point for anyone expecting a signed, click-and-run app.

Where this sits next to other options on this site

This site has also covered clipify, a Claude Code skill that does a related job from the terminal for anyone already living inside Claude Code, and short-video-generator-ai, a hosted tool that trades local processing for zero setup. Shortcast's angle is distinct from both: it's a standalone GUI app, not a coding-agent skill, and it processes everything on the Mac itself rather than a remote server. The tradeoff is real hardware requirements (Apple Silicon, enough RAM to hold a 7 GB model, patience on older M1 chips) in exchange for not handing raw footage to anyone's cloud during editing. Anyone choosing between the three is really choosing between three different constraints: terminal comfort, hosted convenience, or local control.

The traction is genuinely thin

This is the part worth stating plainly rather than talking around. As of today, the repository sits at 361 stars and 56 forks under an Apache 2.0 license. It was created in late May 2026 and hasn't seen a code push since early June, roughly two months of no visible development activity as of this writing. That's not a project with a proven track record or a fast-growing community behind it; it's a single developer's well-documented, technically legitimate build that hasn't yet accumulated much usage signal. The README itself is unusually candid about limitations too: one video at a time, no batch processing, and processing time on an M1 Pro that runs to several minutes for a two-minute clip.

Who this is actually for

Shortcast makes sense for a Mac-owning creator who wants the short-generator workflow but doesn't want raw footage leaving their machine during editing, and who doesn't mind installing an unsigned app or running a one-line Terminal fix to open it. It doesn't make sense as a first pick for anyone without Apple Silicon, anyone who wants a mature, actively maintained project with a real user base behind it, or anyone who needs to batch-process more than one video at a time. The engineering is sound and honestly documented. The traction just isn't there yet, and that's worth knowing before installing it, not after.