Does searching my footage upload it anywhere?
It depends on the tool, and the honest answer is that you cannot tell from the marketing page - you have to check. Three architectures are sold under the same words: full upload, cloud analysis of extracted frames, and on-device analysis where nothing but the app itself ever touches the network.
Published September 10, 2026 · Field notes
Three different architectures are sold with the same vocabulary, and the words "private", "secure" and even "local" get used for all three. The only way to know which one you are buying is to look at where the analysis happens, because that determines what has to leave your machine.
The three architectures
1. Full upload
Your media files are transferred to a service, which stores them and analyses them there. Search then runs against their copy. This is how most collaboration-first platforms work, and it is a reasonable trade when the whole point is that other people need the footage too. It is not compatible with a multi-terabyte archive on upload time alone, and it is often not compatible with client NDAs.
2. Cloud analysis of extracted frames or audio
The full video stays put; the app extracts frames, thumbnails or an audio track and sends those for analysis. Vendors sometimes describe this as "your footage never leaves your computer", which is true of the file and false of the content - a frame is a picture of the thing you signed an NDA about. This is the category worth asking direct questions about.
3. On-device analysis
The model runs on your CPU, GPU or neural engine. The media, the extracted frames and the resulting index all stay on the machine. Typically the model weights are downloaded once, and after that the feature works with the network off. Adobe's Media Intelligence in Premiere Pro works this way, as does Blackmagic's IntelliSearch in Resolve 21, and so does Reelary.
How to verify it yourself in five minutes
Vendor claims are checkable, and the check is cheap:
- Install the tool and let it finish any first-run model download.
- Turn off Wi-Fi and unplug Ethernet.
- Point it at a small folder and index it. Then search.
- If both complete offline, the analysis is local. If either stalls, it is not.
For a stricter answer, watch the traffic rather than the UI. On macOS, Little Snitch or
lsof -i -P will show what the process is connecting to; on Windows, Resource Monitor's Network
tab or Fiddler will. What you are looking for is sustained outbound volume during indexing, which is the
signature of frames being shipped. A few kilobytes on launch is a licence or update check; a hundred
megabytes while indexing is your footage.
Repeat the test after major updates. Architecture changes between versions, and a feature that was local in one release can gain a cloud path in the next.
What Reelary sends
Being specific, because a vague answer here is worth nothing:
- Media, frames, thumbnails, transcripts and vectors: never sent. Indexing, search, captioning and export all run on the machine. The vector database is a local file.
- Model weights are downloaded once from our host when you choose a model - roughly 405MB for the smaller vision model and 815MB for the default one, plus optional extras such as the 148MB speech-to-text model. That is a download, not an upload.
- Licence checks. The free tier does not need one. A Pro licence key entered by hand contacts our server about once an hour with a machine fingerprint to renew a short-lived signed token, and works offline for days on that token. On macOS, purchases made through the App Store are validated by StoreKit locally and do not contact us at all.
You can verify all of this with the offline test above: pull the network, index a folder, search it, export an MP4. All four work.
What NDA and client work actually require
Most agreements are about custody, not encryption: who holds the material, where it is stored, who may access it, and how it is destroyed. A cloud service holds a decrypted copy in order to analyse it, which is a disclosure to a third party whether or not anyone looks at it. That is the clause people fall foul of, and it is why "we are SOC 2 compliant" answers a different question than the one your client asked.
On-device analysis sidesteps the question entirely: there is no third party, so there is nothing to disclose, sub-process or delete.
When on-device is the wrong choice.
Local analysis costs you time and machine capability. If your footage needs to be reviewed by a distributed team, a cloud platform is the correct tool and privacy is not the deciding axis. If you need the strongest available speech and face recognition across a broadcast archive, large cloud models are still ahead of what runs on a laptop. And if your machine is old, a local index is slow to build - that cost is real, it is paid once per library, and no marketing page should pretend otherwise.
Questions people ask about this
How can I tell whether a video search tool uploads my footage?
Disconnect the machine from the network and try to index and search. A tool that analyses on device will complete both; a tool that analyses in the cloud will fail or queue. Do the test after any first-run model download, and repeat it after an update, because the answer can change between versions.
Does on-device analysis mean nothing leaves my machine at all?
Not automatically. An app can analyse footage locally and still send telemetry, licence checks or crash reports. Those are usually small and unrelated to your media, but if you work under an NDA the distinction to verify is whether media, extracted frames, thumbnails or transcripts leave - not whether the process makes any network request at all.
Is uploading footage to a cloud tool a problem if it is encrypted in transit?
Transit encryption is not the issue for most client work. The obligations in a typical NDA concern who holds the material and where it is stored, and a cloud tool holds a decrypted copy in order to analyse it. Check the contract you signed rather than the vendor security page.