Know the boundary before you hand media to an agent.

Voidscape is a local-first CLI and agent skill for turning a local video/audio file or supported public media URL into frames, a timestamped transcript, and a manifest an agent can cite.

You need Python 3.10+, ffmpeg, and ffprobe. yt-dlp is needed for URLs. Local speech transcription additionally uses faster-whisper. Run doctor to see what is ready.

Install the CLI with uv tool install https://github.com/RikepilB/void-scape/archive/refs/heads/main.zip, then run voidscape init. No repository clone is required; init adds the bundled agent skill and checks local media tools.

The CLI can create evidence without a subscription. An agent is what reads that evidence into a grounded answer or Markdown note. Standalone semantic report generation is a future optional direction, not a current CLI claim.

inspect discovers source facts. preview exposes cost, privacy, and dependencies. read prepares the approved evidence.

A work folder contains selected frames/, a transcript.txt when audio is selected, and manifest.json mapping frames to source timestamps.

Yes. The agent should ground answers in the transcript and frames and cite the source timeline as [MM:SS]. Captions and transcripts can contain recognition errors, so important claims should be checked against the evidence.

Not today. The CLI writes evidence. The installed agent skill instructs an agent to create a short TL;DR and timestamped beats, and to save an agent-authored Markdown result when a local Library is configured.

Not by default. Local files, sidecar subtitles, and local transcription stay on your machine. A cloud transcription route requires a current --allow-cloud approval.

It estimates transcription cost and API-equivalent agent-token cost, identifies the dominant driver, shows the backend chain, and tells you whether a cloud approval, model download, or local dependency is needed.

Cloud approval controls sending audio to a selected provider. Model-download approval controls acquiring a local speech model. They are separate decisions and one does not imply the other.

Only in environment variables that you set. Voidscape does not scan browser storage, credentials, or .env files, and its workspace preferences store paths and defaults rather than API keys.

Local videos, recordings, and audio are supported. Public YouTube and other platform URLs work through yt-dlp when the platform allows access. Platform behavior can change.

For media you are permitted to access, you can provide a user-exported Netscape-format cookies.txt path through READ_VIDEO_YTDLP_COOKIES. Keep that file private and outside the repository.

No. It never discovers or extracts browser credentials. A cookie export is optional, local, and chosen by the user for a specific source.

A user-observed repository workflow can help select a permitted saved Reel in a signed-in Chrome session, then the normal media flow handles a confirmed URL. It is not an installed generic Instagram command. Any unsave action needs a separate explicit confirmation after evidence is safely recorded.

Your browser may be signed in while the CLI is anonymous. Expired sessions, VPNs, regional restrictions, platform extractor changes, and missing site permissions can also cause failures. Start with a public URL and read the authenticated-source guide when needed.

Yes. Ask the agent after Voidscape prepares the evidence. The answer should be direct, grounded in the selected artifacts, and include timestamps for factual claims.

Yes. When a workspace Library is configured, the agent can save an authored Markdown note there. The CLI does not claim semantic authorship of that note.

It can draft an email based on evidence and timestamps. Sending email is a separate external action that must remain visible to you and require your explicit approval.

No scheduler or unattended worker ships with Voidscape. You can call the non-interactive CLI from your own tools, but your automation must preserve every cloud and model-download approval gate.

Choose a local transcription backend such as faster-whisper, or preview a cloud backend and explicitly approve it. Captions-only mode correctly stops when a source has none.

Run doctor. It reports Python, FFmpeg, FFprobe, yt-dlp, workspace status, and optional local speech support. Install only the missing tool you have chosen to use.

Yes for local HTML, Markdown, text, RSS/Atom, and local image carousels. Public article URLs require explicit fetch approval. Subscriber-only Substack pages, X or LinkedIn threads, and broader saved-content adapters remain separate browser or adapter work.

Direct public media URLs may work when their platform allows yt-dlp access. Dedicated platform adapters, collection workflows, and browser-extension automation need separate product, platform, and safety work before shipping.

Yes, once you no longer need the frames, transcript, or manifest for follow-up questions. Keep any Markdown note you want to preserve. Delete only folders you recognize and control.