Independent project · video-analysis CLI
Give an agent evidence from video, not a vague summary.
Open Vision turns video into timestamped visual and audio evidence, then returns structured output that an agent can use.

THE QUESTION
The project explores a practical question: how can an agent inspect a long video and return the specific moments that support an answer?
01 / WHAT IT DOES
The system
- A command-line workflow for analyzing video and returning a searchable timeline.
- Visual-frame and audio-transcript evidence with timestamps, exposed as JSON or JSONL.
- A local SQLite index for storing and retrieving timeline data.
02 / ENGINEERING CHOICES
What shaped it
01
Keep the default path local-first
Local media and a local index make inspection easier without requiring a hosted application.
02
Return evidence as data
Machine-readable output lets an agent or another tool consume results without scraping a prose summary.
03
Keep time attached to each signal
Frames and transcript segments remain tied to positions in the source video, so a result can point back to a moment.
03 / SCOPE AND LIMITS
What this does not prove
- Local-first is not local-only. Hosted URLs and remote model providers can involve network services, depending on configuration.
- Timestamped extraction gives an agent evidence to inspect; it does not prove that the agent interpreted that evidence correctly.