Voidscape

Turn the media you keep into local, timestamped evidence before you pay or upload.

Comment

Why I built Voidscape

Saved videos, voice notes, demos, and recordings can become a pile of things I mean to return to. An agent can summarize text and images, but it cannot inspect a video until someone turns that media into evidence it can read.

Voidscape is a local-first CLI and agent skill that turns local recordings, voice material, and supported public video URLs into selected frames, timestamped text, and a manifest that maps evidence back to the source. An agent can then answer with [MM:SS] citations instead of guessing from a title, thumbnail, or prompt.

Transcription is one channel, not the finished product. Voidscape prepares the whole evidence package for an agent: visual frames, source-timeline text, and a machine-readable manifest. If a specialist transcription tool has already produced a matching SRT, VTT, or text file, Voidscape can reuse it locally as a sidecar instead of generating it again.

How it works

  1. Inspect source facts: duration, audio, captions, local sidecars, and likely scope.
  2. Preview cost, dependencies, model-download state, agent-token estimate, and the cloud-privacy boundary.
  3. Read only the approved scope into frames, transcript, and manifest.

The default path stays local. A cloud transcription route is rejected unless explicitly approved with --allow-cloud; a first-time local Whisper model download has its own separate approval. An API key, browser session, or prior run is never treated as consent.

Available now

  • Local recordings, demos, meetings, screen captures, voice material, and supported public video URLs.
  • Visual, audio-only, or combined reads, including scoped time windows with source-timeline timestamps.
  • Matching SRT, VTT, and plain-text sidecars reused locally without retranscription.
  • Frames, timestamped text, manifests, cost previews, and grounded agent answers.
  • Guided inspect, preview, read, customize, and doctor commands.
  • A machine-readable envelope protocol, discovery manifest, stable exit codes, and retryability metadata for agents.
  • A repository-only Instagram URL validation/dedup helper, clearly labeled as not an installed command.

Coming soon — planned, not shipped

  • Private YouTube queue capture through a user-owned playlist.
  • Images, carousels, articles, RSS, and more source adapters.
  • Broader agent/MCP packaging beyond the current skill.
  • A hosted service, browser extension, and unattended workflows remain exploration only.

Built with Codex and GPT-5.6

Codex helped audit and harden the existing engine, implement the guided Voidscape layer, build regression tests, and exercise the actual inspect → preview → consent → read flow. GPT-5.6 is the target agent model for the prepared evidence; Voidscape estimates its 32×32 vision patches so the API-equivalent cost is visible before a run.

The human choices stayed mine: local-first privacy, separate consent for cloud and model downloads, the personal-media use cases, and the decision not to claim unbuilt adapters as product features.

Try it

Supported: Windows PowerShell, macOS/Linux Bash, Python 3.10+, and FFmpeg/FFprobe; yt-dlp is needed only for URLs. No key, account, or copyrighted media is required for this reproducible demo:

.\scripts\install-skill.ps1
python scripts/create-demo-fixture.py
python skill/scripts/voidscape.py inspect samples/build-week-demo.mp4
python skill/scripts/voidscape.py preview samples/build-week-demo.mp4 --tier both --backend captions
python skill/scripts/voidscape.py read samples/build-week-demo.mp4 --tier both --backend captions --workdir samples/build-week-output

Open samples/build-week-output/manifest.json, its frames, and transcript.txt, then ask an agent for a summary with [MM:SS] citations.

Built With

Share this project:

Updates