Cascade

Cascade turns uploaded videos into searchable clips and provides project management, research, caption styling, and a nondestructive video editor. The current website uses FastAPI and Clerk authentication. Generation uses the existing clip agent, Deepgram transcription, configurable AI provider, and parallel video rendering. Editor chat independently uses OpenAI.

Start from the repository root:

python3 -m pip install -r requirements.txt
cp .env.example .env  # Only when creating a new configuration; keep an existing .env.
uvicorn app1:app --reload --host 127.0.0.1 --port 8000

Install FFmpeg/FFprobe separately. See setup and configuration, editor behavior and browser checks, and multi-clip investigation, and project groups and follow-ups.

app1.py                  FastAPI entry point, authentication, upload/generation
backend/                 Project/editor services and feature implementations
web/                     One HTML template (index.html), scripts, and styles
tests/                   Regression and isolated browser checks
  manual/                Explicit opt-in model/render smoke checks
docs/                    Setup, editor, investigation, and cleanup notes
video_jobs/              Private uploads, jobs, projects, drafts, and exports

Run offline regression checks from the repository root:

python3 -m unittest discover -s tests -p 'test_*.py' -q
node tests/test_editor_state.cjs
node tests/test_editor_flow.cjs
node tests/test_canvas_branches.cjs

Some media tests require FFmpeg with the subtitles filter (libass). On Homebrew, install ffmpeg-full and put its bin directory first in PATH. Tests mock paid providers; scripts under tests/manual/ are opt-in and may use configured external services. Do not run them as part of an offline test suite.

Keep .env, video_jobs/, outputs/, and other saved media private. They are not disposable build output. All seven page views live in web/index.html; existing page URLs remain valid. The legacy publishable-key fallback was migrated there without changing authentication. See frontend routing and verification.

Pexels B-roll is available in the editor; see B-roll setup and workflow.

Fill to Frame automatically uses local audio-visual Follow Speaker tracking when a crop is needed. Download its public models once with python3 scripts/setup_reframe.py; see Auto Reframe for setup, honest tracking limitations, and verification. Export selection is in the editor’s collapsible Session/Wave sidebar; Exports mirrors that hierarchy.

See the editor and Exports delivery update for grouped downloads, compact/expanded tracks, Research Agent preferences, B-roll dragging, and current verification limits.