You remember the shot. Your drive only knows it as C0042.MP4.
Vidi indexes your footage on your own machine and finds the moment from a plain description. It runs offline once the model is installed, never alters a byte of the original, and exports a timeline that opens in Premiere and Resolve.
- Windows 11 x64
- macOS 14+ Apple Silicon
The problem
The take exists. The problem is remembering which drive it's on.
You remember the shot: woman crossing the wet street, backlit, red umbrella. Some day downtown. The file is called A012_C007_0421XY.mov, and logging stopped on day three of the shoot. So the ritual starts: mount the drive, open folder after folder, scrub the playhead, ask the group chat if anyone remembers. An hour later you give up and reshoot something you already had. The obvious fix is a cloud system, until you price out pushing 12 TB through an office connection, add the per-seat fee, and remember that half your contracts forbid the footage leaving your network. The archive stays on the drive. And stays invisible.
Describe the scene, get the file and the moment
Type "people walking on the beach at sunset". Clips come back ranked by relevance, grouped by video, with a thumbnail, the full path and the position inside the file. One click previews it, another reveals it in Explorer. Too few results? Lower the threshold. It reorders instantly, nothing gets reindexed.
Send the selects to your editor without copying a gigabyte
Mark the clips, pick the order, set your handles, export. You get an .otio pointing at the original files right where they are. It opens in Premiere Pro 25.6+ and Resolve 20.3+. Older Premiere gets FCP7 XML from the same manifest. A preflight first lists offline media, files outside the authorised root, and handles that ran past the end. You find out in Vidi, not in front of the client.
Land on the right frame, including 23.976 and 29.97 drop-frame
This is the part that takes work and nobody sees. Vidi never uses a float as a source of time: every clip carries the original integer PTS, the time base as a fraction, the start timecode, the drop-frame flag and rotation. Conversion happens only when writing the editor's file. You don't have to take my word for it. Open the .otio in a text editor.
Pull the network cable and keep working
The network is used once: downloading the model pack, with its checksum and origin shown on screen. An offline machine imports the same pack as a file. After that, cataloguing, searching and exporting never touch the network. There is no account, no login, no upload. Easy test: block the executable in your firewall and work all day.
Index overnight or while you edit: your call
Three resource profiles with the trade-off written on screen, plus an option to force CPU and leave the whole GPU to your editor. The app measures what the machine can actually take before the first batch and ramps up gradually; when memory gets tight it backs off instead of killing the job. Pause, reboot, resume: it picks up from the last intact checkpoint.
Your footage stays byte for byte as it was
Read-only source media is an architectural rule, not a setting: no code path alters, moves or deletes your files. The index, thumbnails and database live in the app's own directory, never inside your media folder. An auditor takes a SHA-256 snapshot of the tree before and after every test run and fails if a single byte changed. Deleting the whole catalogue costs you nothing.
Frequently asked
- Does this force me to upload everything?
- No. There is no upload feature and no user account. Processing happens in a local service that only talks to itself, over authenticated loopback. The network is used at exactly one moment: fetching or updating the model pack, with consent and a checksum.
- Will it find anything I ask for?
- No, and you should know that now. This version indexes visual content only, by sampling. Speech, on-screen text, music and motion are not included. If the scene only exists because someone said a sentence, Vidi won't find it. Very short events can slip past the sampling too. Those signal sources are in the backlog, after 1.0, with no date promised.
- I'm on a Mac. Can I get in?
- Not today. The beta is Windows 11 x64. macOS on Apple Silicon is a declared target and the code was written for both platforms from day one, but the Mac side has zero execution: no benchmark run, no signed build, no test on a real machine. I'd rather say that now than put you on a list pretending a date exists.
- If it works, why is there no download button?
- "It runs on my machine" is not "it runs on yours". The full pipeline is implemented and covered by tests, the packaged app passes startup tests on CUDA and CPU, and two Windows installers exist. What is missing isn't code: it's a signing certificate, an install verified on a genuinely clean machine, the pinned release FFmpeg build, and all of macOS. Windows treats an unsigned installer as suspicious, and it's right to. So the beta is hand-picked.
- Tauri v2
- Rust
- React 19
- Python 3.12
- FastAPI
- SQLite
- LanceDB
- FFmpeg
- CLIP
Before I wrote a line of Vidi, I was a logger. I worked location shoots in the United States, organising the raw footage and handing it over to editors who were in another country. On the broadcast of Nubank's IPO, live from the New York Stock Exchange, my job was making sure whoever cut it could find the right take without ever having been there. Before that, five years at a production house, running the studio's technical side and the infrastructure at once. Vidi is the tool I wanted in both jobs.
João Pedro Mantovani · Engenharia de Computação · Instituto Mauá de Tecnologia