Is this just a chat wrapper over keyword search?

No — the difference is that it plans. A question naming a person and a topic becomes a face search intersected with a transcript search, run as separate calls and combined. You can watch it happen in the step list, including the result count each call returned.

Why does it show every tool call instead of just the answer?

Because an answer you can't audit is not much use when you're about to cut with it. Seeing that the face search returned 20 and the topic filter cut it to 3 tells you whether to trust the result or loosen the query — and it makes a wrong answer diagnosable rather than mysterious.

Does the agent run a local model, or does it call out to the cloud?

Both, and you choose. Point the agent at a local model over Ollama or LM Studio and nothing ever leaves your Mac. Prefer a hosted model (Claude, OpenAI) and only your question and the search results go over the network — never your footage or the raw index. The default is the local route.

Does asking it a question send my footage anywhere?

No. The index and the matching stay on your Mac either way. With a hosted model the only thing that crosses the network is the text of your question and the result summaries the agent reasons over — that's the point of the local-first design: you can put footage under NDA through it without a clause to worry about.

What happens when it genuinely can't find anything?

It tells you, and the step list shows where the funnel emptied out. That's more useful than a confident wrong clip — if the person was never labelled, or the topic was never said on camera, no amount of rephrasing will conjure it.

Core feature

Agent

Ask for the moment in plain English — the agent works out how to find it.

One question, a visible plan, the tools it picked, and the scenes that came back.

Every other way of searching asks you to know which index to reach for — words, faces, color, camera data. The agent removes that step. You describe the moment the way you'd describe it to an editor sitting next to you, and it decides which tools to run, in what order, narrowing as it goes. It shows its working the whole time: each tool it calls, what it passed, how many results came back, and how long it took.

Get Orcah Studio for Mac

Requires an Apple Silicon Mac, macOS 14.0 or later

$249 once · lifetime license · 30-day refund, no questions asked

Getting started with Agent

1.
Point Orcah at a folder
The agent can only reason over what's indexed. Once a library is in, every layer — transcript, faces, scenes, color, telemetry, camera metadata — becomes a tool it can reach for.
2.
Ask the way you'd ask a person
“Find me videos of @Ilias talking about imposter syndrome.” Name people with @, and scope to a project or the whole library from the composer.
3.
Watch the plan, then send the result
Steps appear as they run and resolve to result counts. When the answer lands, tick the clips you want and send them straight to Resolve or Final Cut (Premiere Pro by request access).

What it can do

Multi-step search: it chains tools — narrowing by person, then topic, then what's on screen — instead of running one query.
Visible reasoning: every tool call is shown with its arguments, result count and elapsed time, so you can see why a clip surfaced.
@mentions: name a labelled person directly in the question to pin the search to them.
Scoped runs: point it at one project or the entire library from the composer.
Straight to the timeline: approve the clips you want and send them across linked in place, with no re-render.
Your call where it thinks: run the reasoning model locally (Ollama, LM Studio) so nothing leaves the Mac, or point it at a hosted model — the index and footage stay on your Mac either way.

Pro Tip: when a run comes back thin, read the steps rather than rewording the question — the result counts show you exactly which filter collapsed, and that usually tells you what to relax.

Questions worth asking

Is this just a chat wrapper over keyword search?: No — the difference is that it plans. A question naming a person and a topic becomes a face search intersected with a transcript search, run as separate calls and combined. You can watch it happen in the step list, including the result count each call returned.
Why does it show every tool call instead of just the answer?: Because an answer you can't audit is not much use when you're about to cut with it. Seeing that the face search returned 20 and the topic filter cut it to 3 tells you whether to trust the result or loosen the query — and it makes a wrong answer diagnosable rather than mysterious.
Does the agent run a local model, or does it call out to the cloud?: Both, and you choose. Point the agent at a local model over Ollama or LM Studio and nothing ever leaves your Mac. Prefer a hosted model (Claude, OpenAI) and only your question and the search results go over the network — never your footage or the raw index. The default is the local route.
Does asking it a question send my footage anywhere?: No. The index and the matching stay on your Mac either way. With a hosted model the only thing that crosses the network is the text of your question and the result summaries the agent reasons over — that's the point of the local-first design: you can put footage under NDA through it without a clause to worry about.
What happens when it genuinely can't find anything?: It tells you, and the step list shows where the funnel emptied out. That's more useful than a confident wrong clip — if the person was never labelled, or the topic was never said on camera, no amount of rephrasing will conjure it.

Find the moment. Not the file.

Download immediately, models included. Your footage stays on your Mac.