Prerequisites & system requirements
Agent28 ships as three separate builds: macOS on Apple Silicon, macOS on Intel, and Windows x64. Each section below lists what that platform needs, what is bundled, and what downloads or installs on first use. Requirements that apply everywhere come first.
Which download do I need?
| Platform | Download | How to tell |
|---|---|---|
| Mac with Apple Silicon (M1–M4) | Agent28-<version>-arm64.dmg | Apple menu → About This Mac shows “Apple M1 / M2 / M3 / M4…” |
| Mac with an Intel processor | Agent28-<version>-x64.dmg | About This Mac shows “Intel Core …” under Processor |
| Windows PC (64-bit) | Agent28 Setup <version>.exe | Settings → System → About shows “64-bit operating system, x64-based processor” |
Important: The two Mac builds are not interchangeable. The Apple Silicon build refuses to open on an Intel Mac, and the Intel build on an Apple Silicon Mac only runs through Rosetta with no Metal acceleration — always pick the one that matches your chip.
All platforms
- Account. Sign in with your Agent28 account. Agent28 AI (in-app chat) needs Plus, Pro, or Studio; on Free, in-app chat uses your own API keys. ChatGPT and Claude work on every plan.
- Network. Sign-in opens
app.agent28.comin your browser and managed AI talks toapi.agent28.com. Local models download once, on first use, fromhuggingface.coandstorage.googleapis.com. Rendering, transcription, scene indexing, and person matting all run on your machine and work offline after the models are cached. - Disk. About 1.5 GB for the app itself, plus models downloaded on first use — Fast transcription 141 MB, Whisper about 1.6 GB, Best about 2.6 GB, plus small face-detection and matting models — and project caches (proxies, transcripts, scene indexes). Keep at least 10 GB free.
- Memory. 8 GB minimum. 16 GB recommended if you use person matte, scene indexing, and the Best transcription engine on the same project.
- GPU.Any GPU works. Person matte and background removal run through WebGPU (Metal on macOS, Direct3D 12 on Windows) and fall back to the CPU automatically if the GPU is unavailable. Video decoding uses the operating system's hardware decoder.
- Display. The window opens at 1440 × 900 and can be shrunk to 1024 × 600. Smaller screens are not supported.
macOS · Apple Silicon
| Requirement | Details |
|---|---|
| Hardware | Any Mac with an Apple M1, M2, M3, or M4 chip. |
| Download | Agent28-<version>-arm64.dmg. Rosetta is not required. |
| macOS to run the app | macOS 11 Big Sur or later. |
| macOS for the local AI sidecar | macOS 13 Ventura or later. The sidecar powers scene indexing, footage search, beat detection, object tracking, and web-footage import. On macOS 11–12 the app runs without those features. |
| Install | The build is notarized by Apple. Open the DMG, drag Agent28 to Applications, then launch from Applications (the first launch may take a few seconds while macOS verifies the app). |
| Bundled | Fast transcription (Whisper base with Metal acceleration) and the local AI sidecar. Nothing to install. |
| Optional | Python 3.10 or newer for the Best and Whisper transcription engines — see Transcription engines below. |
macOS · Intel
| Requirement | Details |
|---|---|
| Hardware | An Intel Mac whose CPU supports AVX2 — Haswell (2013) or newer. That is every Intel Mac that can run macOS 11 except the Mac Pro (Late 2013), whose Xeon lacks AVX2; on that machine the bundled Fast engine is unavailable and the other engines still work. |
| Download | Agent28-<version>-x64.dmg. |
| macOS to run the app | macOS 11 Big Sur or later. |
| macOS for the local AI sidecar | macOS 13 Ventura or later, same as Apple Silicon. On macOS 11–12 the app runs without scene indexing, footage search, beat detection, object tracking, or web-footage import. |
| Install | Notarized by Apple. Open the DMG, drag Agent28 to Applications, launch. |
| Bundled | Fast transcription (Whisper base, CPU) and the local AI sidecar (Intel build). Nothing to install. |
| Performance | Intel Macs have no Neural Engine, so AI features are noticeably slower than on Apple Silicon. Person matte uses the Intel or AMD GPU through WebGPU where available, otherwise the CPU. The Best engine runs on the CPU only. |
| Optional | Python 3.10 or newer for the Best and Whisper engines — same steps as Apple Silicon. |
Windows · x64
| Requirement | Details |
|---|---|
| Operating system | Windows 10 or Windows 11, 64-bit. Windows on ARM (Snapdragon PCs) is not supported: the x64 build only runs there through emulation and is untested. |
| Hardware | An x64 CPU with AVX2 — Intel from 2013 (Haswell) or AMD from 2015 onward. The bundled Fast transcription engine needs AVX2; everything else runs on any x64 CPU. |
| Download | Agent28 Setup <version>.exe — a standard installer that lets you choose the folder and adds Desktop and Start Menu shortcuts. |
| SmartScreen | The Windows installer is not yet code-signed, so SmartScreen can show “Windows protected your PC”. Click More info → Run anyway. The download itself comes only from agent28.com. |
| Bundled | Fast transcription (Whisper base), Whisper transcription (faster-whisper), and the local AI sidecar. No Python is needed for any of these. |
| GPU | A DirectX 12 capable GPU enables WebGPU for person matte and background removal; without one those features run on the CPU. Video decode uses Media Foundation hardware decoding when available. |
| Optional | Python 3.10 or newer for the Best engine. To run it on the GPU, install onnxruntime-directml (any DirectX 12 GPU, no CUDA needed) or onnxruntime-gpu (NVIDIA, CUDA) into that Python — never both. |
What works where
| Feature | Mac · Apple Silicon | Mac · Intel | Windows · x64 |
|---|---|---|---|
| Editor, timeline, captions, motion graphics, export | Yes | Yes | Yes |
| Agent28 AI, ChatGPT, Claude, Claude Code | Yes | Yes | Yes |
| Fast transcription (bundled) | Yes · Metal | Yes · CPU, needs AVX2 | Yes · CPU, needs AVX2 |
| Best transcription (Parakeet) | Python 3.10+ | Python 3.10+, CPU only | Python 3.10+, optional GPU |
| Whisper transcription (faster-whisper) | Python 3.10+ | Python 3.10+ | Yes · bundled |
| Scene indexing, footage search, beat detection, object tracking, web-footage import | macOS 13+ | macOS 13+ | Yes |
| Person matte & background removal | Yes · GPU or CPU | Yes · GPU or CPU | Yes · DirectX 12 GPU or CPU |
Transcription engines
Pick an engine in the captions panel. Fast needs no setup on any platform and downloads a 141 MB model the first time. OpenAI only needs an OpenAI key under Settings → API Keys. The two other local engines use a Python you install; Agent28 finds it automatically in the locations below.
Best engine on macOS (Apple Silicon and Intel)
Install Python 3.10 or newer (for example Homebrew brew install python@3.12, or the installer from python.org), then create the environment Agent28 looks for:
python3 -m venv ~/.motionfly/parakeet-venv
~/.motionfly/parakeet-venv/bin/pip install onnxruntime numpy huggingface_hubThen choose Best (local) in the captions panel. The first transcription downloads about 2.6 GB of model files into the Hugging Face cache. On Intel Macs pip automatically selects the last onnxruntime release with Intel wheels; the engine runs on the CPU.
Tip: Prefer a different Python? Set the PARAKEET_PYTHON environment variable to any interpreter that can import onnxruntime.
Best engine on Windows
Install Python 3.10 or newer from python.org or the Microsoft Store, then in PowerShell or Command Prompt:
python -m venv %USERPROFILE%\.motionfly\parakeet-venv
%USERPROFILE%\.motionfly\parakeet-venv\Scripts\pip install onnxruntime numpy huggingface_hubFor GPU acceleration replace onnxruntime with onnxruntime-directml (any DirectX 12 GPU — NVIDIA, AMD, or Intel — with no driver matching) or onnxruntime-gpu (NVIDIA with a CUDA-capable driver). Install exactly one of the three; they conflict.
Whisper engine on macOS
On Windows this engine is bundled. On macOS it runs through the python3on your PATH (Homebrew's Python is found automatically), which needs the faster-whisper package:
python3 -m pip install faster-whisperIf your Python refuses system-wide installs, create a virtual environment, install faster-whisper there, and set TRANSCRIPTION_PYTHON to that environment's bin/python. The first transcription downloads the large-v3-turbo model (about 1.6 GB).
AI apps (ChatGPT, Claude, Claude Code)
- ChatGPT desktop — macOS or Windows, with a plan that supports apps and connectors.
- Claude desktop — macOS or Windows. Quit and reopen it once after Agent28 links it.
- Claude Code — the terminal tool on macOS or Windows, signed in.
- The AI app must be on the same computer as Agent28, and Agent28 must stay open while it edits. Enable it under Settings → AI apps.
Troubleshooting
- “Agent28 can't be opened” or “Bad CPU type” on a Mac. You installed the build for the other chip. Check About This Mac and download the matching DMG from the table at the top.
- macOS says the app is damaged or from an unidentified developer. Both Mac builds are notarized by Apple, so this means the download was corrupted. Delete it and download again from agent28.com.
- Scene indexing or footage search never finishes on a Mac. Those features need macOS 13 Ventura or later. On macOS 11–12 the rest of the app works normally.
- Fast transcription fails immediately on Windows or an Intel Mac. The CPU lacks AVX2. Use the Best or Whisper engine, or OpenAI.
- Best or Whisper engine says Python was not found. Follow the engine steps above; on macOS make sure
python3 --versionreports 3.10 or newer in a fresh Terminal window. - Windows SmartScreen blocks the installer. Click More info → Run anyway.