RTSP Snapshot

Use to grab stills or short clips from IP security cameras (RTSP, HTTP HLS) so Kate can report what is happening in a scene or store evidence. All output stays under a sandbox dir (RTSP_SNAPSHOT_OUTPUT_ROOT, default $TMPDIR).

Use when

• "take a snapshot from the garage camera"
• "record 30s from the doorbell now"
• "is there motion? capture and send me a snapshot"

Do not use when

• The camera is not reachable over LAN/public RTSP/HTTP URL
• Necesitas streaming continuo o WebRTC
• Clips > 5 minutes (hard cap by design)

Tools

status

ffmpeg bin + sandbox + limits.

snapshot { url, output_path, transport?, width? }

• url — rtsp://user:pass@host/stream, rtsps://, http://, https:// (NO file://, NO concat:)
• output_path — must stay under the sandbox root; .jpg recommended
• transport — tcp (default, estable sobre NAT) o udp
• width — resize manteniendo aspecto

Returns {url, output_path, bytes, transport}.

clip { url, output_path, duration_secs, transport? }

• duration_secs 1..=300
• Uses stream copy (no re-encode) — very fast

Returns {url, output_path, bytes, duration_secs, transport}.

Execution guidance

• Prefer TCP transport to avoid packet loss over Wi-Fi/VPN
• Credentials are inline in URL (rtsp://user:pass@host/...); store them in 1Password and read them with read_secret when reveal is enabled
• Output path outside sandbox → -32034 IoError; adjust operator config
• Unreachable camera → -32032 NonZeroExit with ffmpeg stderr included
• Motion detection is out of scope here — chain with a vision-enabled pipeline

Pipelines

Snapshot + vision

1. rtsp-snapshot.snapshot { url, output_path: "/sandbox/cam1.jpg" }
2. (futuro) vision-lm.describe { path: "/sandbox/cam1.jpg" }

Clip + transcribe

1. rtsp-snapshot.clip { url, output_path: "/sandbox/ring.mp4", duration_secs: 30 }
2. video-frames.extract_audio { path, output_path: "/sandbox/ring.wav", codec:"wav", mono:true, sample_rate:16000 }
3. openai-whisper.transcribe_file { file_path: "/sandbox/ring.wav" }