A local AI media library for small video teams.
arkiv ingests your footage, transcribes what is said and describes what is shown, and indexes both on your own machine.
Then you find shots by asking for them, in Chinese, Japanese or English.
It follows a film-industry DIT workflow: offload, name, verify, index.
Big crews have a DIT for this. On a small team, the work starts when the shoot ends.
arkiv batch-indexes your footage with local models and a fixed pipeline: no cloud service, nothing to pay per hour.
Before installing, read the requirements · 中文版
Every file is hash-verified (xxh3), interrupted copies resume, and each destination gets an ASC MHL v2 manifest. It never deletes the source card.
A local vision model describes each scene and tags it, so footage with no dialogue is just as searchable as interviews.
It matches the meaning of scene descriptions and transcripts, not filenames. Add camera, tag or language filters on the same line.
GOOD, REV and N·G stay on the footage; Smart Pools split it into to-review, usable and rejected.
Seek to the exact spot a line is spoken. Export SRT, VTT or plain text. Whisper hallucinations are filtered in four layers.
Mark with i / o or drag; marks snap to real frame boundaries, 59.94p included, and exports use the range.
Search arkiv inside Resolve and import results into the Media Pool with clip colors (GOOD green, REV yellow, N·G orange), binned by camera. EDL and FCPXML use the camera's own start timecode plus your IN point, so relinking lands on the same frame; 29.97 / 59.94 use drop-frame automatically.
A production-line training film's footage, fully indexed: 1,161 of the clips are dialogue-free B-roll. One RTX 4070, 100% index coverage.
Transcription, vision, indexing and search all run locally. No telemetry, no update check, no online license check. The only automatic internet traffic is downloading public AI models the first time.
MCP lets an AI assistant query arkiv's text layer: scene descriptions, transcripts, timecodes, tags, filenames and camera metadata (GPS included, if recorded). Video and images never pass through it. If the assistant is a cloud model, that text goes to its provider. It is off unless you configure it.
Client commissions, day-to-day studio work and grant-funded productions are all covered. No separate license, no need to tell us.
Videos, timelines, EDL / FCPXML exports and thumbnails carry no restriction from this license.
This applies to your own use too (Additional Terms 1–2). Installations in use before 1.1.0 keep their exemption.
Turning it into a competing media-management tool, offering indexing as a service to third parties, or re-implementing it as a replacement — even for free.
The test is what your client is buying: work you finished with arkiv is ordinary tool use; arkiv's functionality itself (as a product, service, library or plug-in) is competition. Full terms: PolyForm Perimeter 1.0.1.
| Free | Pro | |
|---|---|---|
| All indexing and search | Yes | Yes |
| Commercial use | Yes | Yes |
| Projects | 3 installs from before 1.1.0: unlimited | Unlimited |
| Cross-project search | — installs from before 1.1.0: kept | Yes |
| Price | NT$0 | NT$3,000one-time |
One-time, perpetual, no subscription. Unlimited projects and cross-project search; updates are free while the add-on is maintained.