CDAF
An open-source video description format lets AI agents reuse timestamped text, with local model generation and optional Gemini processing.
CDAF saves reusable text descriptions alongside videos so AI agents don't have to analyze the same footage for every task. It's for developers building video agents and editors that repeatedly search or ask questions about a clip library.
- Footage search: Agents can search plain-text descriptions across a video library to find clips that match a request.
- Editing timestamps: Segment times support cut decisions in workflows such as Remotion and HyperFrames, while timed transcripts provide speech alignment for captions and audio sync.
- Stale description checks: Each description records the video's hash and file size, so conforming tools reject it when the footage changes.
- Agent reuse: The included Claude Code skill teaches agents to check for a description and verify its freshness before analyzing a video; the format is model agnostic, so other agents can read the same files.
The Python library and command-line tool generate, read and validate descriptions. Reading and validation run locally without an API key or dependencies beyond Python's standard library.
CDAF is open source under the MIT license and works on Windows, macOS and Linux. Local model generation doesn't require an API key. The optional Gemini backend sends footage to Google's Gemini Files API using your own key. The Python tool requires Python 3.10 or later.
