Ffmpeg Capture

Ffmpeg Capture

Context: $ARGUMENTS

Quick start

• Record the screen: → Step 1 (list devices) → Step 2 (pick source) → Step 3 (run)
• Record webcam + mic: → Step 1 → Step 3 with webcam subcommand
• Record a region: → Step 3 with --crop (mac/linux) or -offset_x/-offset_y/-video_size (Windows gdigrab)
• Record system audio (mac): install BlackHole first → Step 2 (route via virtual device)
• Just list devices: python3 scripts/capture.py list-devices

When to use

• User wants a screencast, demo recording, or desktop capture to a file.
• User wants to record webcam/microphone/system audio to disk.
• User wants to enumerate which cameras/mics/screens ffmpeg can see.
• User wants low-latency or crash-safe recording (MKV-first, remux to MP4 after).

Do NOT use for: streaming to RTMP/HLS (use ffmpeg-streaming), applying filters/effects (use ffmpeg-video-filter), or converting an existing file (use ffmpeg-transcode).

Step 1 — Detect platform + list devices

Always list devices first. Names and indices are shown in the format ffmpeg expects them.

macOS (avfoundation):

ffmpeg -f avfoundation -list_devices true -i ""

Output has two sections: AVFoundation video devices: (cameras + screens, screen is usually the last one with Capture screen 0) and AVFoundation audio devices:. Indices reset between the two lists — video 1 and audio 0 are unrelated.

Windows (dshow for camera/mic, gdigrab for desktop):

ffmpeg -list_devices true -f dshow -i dummy
ffmpeg -list_options true -f dshow -i video="HD Webcam"

gdigrab needs no enumeration — input name is always desktop (or title=Window Name).

Linux:

# Cameras
v4l2-ctl --list-devices
ffmpeg -f v4l2 -list_formats all -i /dev/video0
# Audio (Pulse)
pactl list sources short
# Audio (ALSA)
arecord -L
# Display (X11)
echo "$DISPLAY"   # usually :0.0 or :1.0

Or run the helper: python3 scripts/capture.py list-devices (auto-detects platform).

Step 2 — Pick source + audio route

Decide:

1. Video source: screen? camera? region? window?
2. Audio route: mic? system audio? both? none?
3. Container: MKV for long/crash-safe sessions, MP4 for quick clips.
4. Encoder: libx264 -preset ultrafast -pix_fmt yuv420p for CPU; platform HW encoder for 60fps+ (see [references/devices.md](references/devices.md)).

System audio notes:

• macOS has no native loopback — install BlackHole (or use Loopback / Rogue Amoeba). Route system output to BlackHole, then use -i ":". To hear it too, build a Multi-Output Device in Audio MIDI Setup.
• Windows: use -f dshow -i audio="Stereo Mix" (if enabled) or VB-Cable.
• Linux: pactl list sources short shows a *.monitor source for each sink — that's system audio.

Step 3 — Run capture with proper encoder

macOS — screen + default mic (most common)

ffmpeg -f avfoundation -framerate 30 -capture_cursor 1 -i "1:0" \
  -c:v libx264 -preset ultrafast -pix_fmt yuv420p -c:a aac -b:a 160k out.mp4

"1:0" = video device index 1 (screen), audio device index 0 (default mic). Replace with your actual indices from Step 1.

macOS — screen only, no audio, 60fps, HW-encoded

ffmpeg -f avfoundation -framerate 60 -capture_cursor 1 -i "1" \
  -c:v h264_videotoolbox -b:v 8M -pix_fmt yuv420p out.mp4

macOS — crop a region from full-screen capture

ffmpeg -f avfoundation -framerate 30 -i "1" \
  -vf "crop=1280:720:100:100" -c:v libx264 -preset ultrafast -pix_fmt yuv420p out.mp4

avfoundation has no native region mode; always full-screen then crop via -vf crop=W:H:X:Y.

macOS — built-in camera + mic

ffmpeg -f avfoundation -framerate 30 -video_size 1280x720 -i "0:0" \
  -c:v libx264 -preset ultrafast -pix_fmt yuv420p -c:a aac out.mp4

macOS — screen + mic + system audio (BlackHole)

Assume BlackHole is audio index 2, mic is index 0:

ffmpeg -f avfoundation -framerate 30 -capture_cursor 1 -i "1:0" \
       -f avfoundation -i ":2" \
       -filter_complex "[0:a][1:a]amix=inputs=2:duration=longest[a]" \
       -map 0:v -map "[a]" -c:v libx264 -preset ultrafast -pix_fmt yuv420p -c:a aac out.mp4

Windows — full desktop

ffmpeg -f gdigrab -framerate 30 -i desktop \
  -c:v libx264 -preset ultrafast -pix_fmt yuv420p out.mp4

Windows — region of desktop

ffmpeg -f gdigrab -framerate 30 -offset_x 100 -offset_y 100 \
  -video_size 1280x720 -i desktop -c:v libx264 -preset ultrafast -pix_fmt yuv420p out.mp4

Windows — specific window

ffmpeg -f gdigrab -framerate 30 -i title="Mozilla Firefox" -c:v libx264 out.mp4

Windows — webcam + mic (dshow)

ffmpeg -f dshow -framerate 30 -video_size 1280x720 \
  -i video="HD Pro Webcam C920":audio="Microphone (Realtek)" \
  -c:v libx264 -preset ultrafast -pix_fmt yuv420p -c:a aac out.mp4

Linux — X11 full display

ffmpeg -f x11grab -framerate 30 -video_size 1920x1080 -i :0.0 \
  -c:v libx264 -preset ultrafast -pix_fmt yuv420p out.mp4

Linux — X11 region

ffmpeg -f x11grab -framerate 30 -video_size 1280x720 -i :0.0+100,100 out.mp4

Linux — webcam (v4l2)

ffmpeg -f v4l2 -input_format mjpeg -video_size 1280x720 -framerate 30 \
  -i /dev/video0 -c:v libx264 -preset ultrafast -pix_fmt yuv420p out.mp4

Linux — audio (Pulse or ALSA)

ffmpeg -f pulse -i default out.wav           # Pulse default mic
ffmpeg -f pulse -i alsa_output.monitor out.wav  # system audio (monitor)
ffmpeg -f alsa -i hw:0 out.wav

Linux — DRM/KMS (Wayland / no X11)

sudo ffmpeg -device /dev/dri/card0 -f kmsgrab -i - \
  -vf 'hwdownload,format=bgr0' -c:v libx264 -preset ultrafast -pix_fmt yuv420p out.mp4

Requires root or sudo setcap cap_sys_admin+ep $(which ffmpeg).

Step 4 — Stop cleanly

• Press q in the ffmpeg terminal. ffmpeg flushes the mux and finalizes the MOOV atom for MP4.
• Do NOT Ctrl+C unless recording MKV. Ctrl+C on MP4 often leaves an unreadable file (no MOOV). MKV is crash-safe and can be remuxed later: ffmpeg -i recording.mkv -c copy recording.mp4.
• For unattended timed captures use -t 60 (stop after 60s) rather than killing the process.
• For long sessions use -f segment -segment_time 600 -reset_timestamps 1 out_%03d.mkv so crashes cost at most one segment.

Available scripts

• scripts/capture.py — cross-platform wrapper: list-devices, screen, webcam, audio-only. Auto-detects the platform and builds the right ffmpeg command. --dry-run prints the command without running it.

Workflow

# 1. What devices exist?
python3 ${CLAUDE_SKILL_DIR}/scripts/capture.py list-devices

# 2. Dry-run to confirm the command
python3 ${CLAUDE_SKILL_DIR}/scripts/capture.py screen --output out.mp4 --fps 30 --dry-run

# 3. Record
python3 ${CLAUDE_SKILL_DIR}/scripts/capture.py screen --output out.mp4 --fps 30 --audio auto

Reference docs

• Read [references/devices.md](references/devices.md) for the full per-platform option tables (avfoundation, gdigrab, dshow, x11grab, kmsgrab, v4l2, alsa, pulse, decklink), virtual-audio devices, TCC/permission reset, crash-safe recording, and bitrate/FPS recommendations.

Gotchas

• macOS permissions: ffmpeg needs Screen Recording AND Microphone permission in System Settings → Privacy & Security. First run prompts the dialog and ffmpeg exits immediately with empty/EAGAIN capture — grant permission to the parent terminal/IDE (Terminal.app, iTerm, VS Code) and re-run. Permission is per-binary-path: reinstalling ffmpeg to a new location re-prompts.
• avfoundation indices are not arbitrary. "1:0" means "video device 1, audio device 0" exactly as listed by -list_devices true. Video and audio lists are numbered independently.
• avfoundation has no system-audio input. Install BlackHole / Loopback and chain it as a second -f avfoundation -i ":N" input, then -filter_complex amix or -map to combine.
• dshow device names need exact case and quoting on Windows. video="HD Webcam" (curly-quote will break). Audio and video are separate arguments in the same -i: video="X":audio="Y".
• -pix_fmt yuv420p is required for libx264 output that plays in QuickTime/Safari/iOS. Leave it off and you get yuv444p, which many players refuse.
• 60fps screen capture saturates libx264-ultrafast on most CPUs. Use h264_videotoolbox (mac), h264_nvenc (NVIDIA), h264_qsv (Intel), h264_amf (AMD) — see ffmpeg-hwaccel.
• gdigrab cursor: -draw_mouse 1 (default on). Some builds ignore it; cursor overlay is unreliable in virtual display setups (RDP).
• x11grab :0.0 must match the running $DISPLAY. On multi-monitor X, use :0.0+X,Y to pick the origin.
• v4l2 -input_format mjpeg is required for most UVC webcams at 720p/1080p 30fps. Raw YUYV falls back to 5–10fps at full res.
• kmsgrab requires root or setcap cap_sys_admin+ep. Outputs DRM prime frames — always chain hwdownload,format=bgr0 before encoding on CPU.
• Pulse source names change between reboots on some systems — pactl list sources short each session.
• Stopping: q in terminal = clean shutdown. Ctrl+C on MP4 = corrupted file. Record MKV for anything over a minute, remux after.
• Audio drift: on long sessions audio can drift several hundred ms. Use -async 1 or record audio to a separate file and mux after.

Examples

Example 1: quick 30-second screen recording on macOS

# List devices
ffmpeg -f avfoundation -list_devices true -i ""
# → video index 1 = "Capture screen 0", audio index 0 = "MacBook Pro Microphone"

ffmpeg -f avfoundation -framerate 30 -capture_cursor 1 -i "1:0" \
  -t 30 -c:v libx264 -preset ultrafast -pix_fmt yuv420p -c:a aac demo.mp4

Example 2: crash-safe 2-hour tutorial recording on Windows

ffmpeg -f gdigrab -framerate 30 -i desktop \
       -f dshow -i audio="Microphone (USB Audio)" \
       -c:v libx264 -preset veryfast -crf 23 -pix_fmt yuv420p -c:a aac \
       -f segment -segment_time 600 -reset_timestamps 1 tutorial_%03d.mkv
# Press q to stop. Then remux:
ffmpeg -f concat -safe 0 -i <(for f in tutorial_*.mkv; do echo "file '$PWD/$f'"; done) -c copy tutorial.mp4

Example 3: webcam-only on Linux for a video call recording

ffmpeg -f v4l2 -input_format mjpeg -video_size 1920x1080 -framerate 30 -i /dev/video0 \
       -f pulse -i default \
       -c:v libx264 -preset ultrafast -pix_fmt yuv420p -c:a aac call.mp4

Troubleshooting

Error: Input/output error or zero bytes (macOS)

Cause: Screen Recording permission not granted to the parent terminal. Solution: System Settings → Privacy & Security → Screen Recording → enable Terminal/iTerm/VS Code. Fully quit and relaunch the terminal (not just re-run). If still failing, reset TCC: tccutil reset ScreenCapture.

Error: Could not find video device with name [...] (Windows dshow)

Cause: Device name mismatch (case, trailing spaces, truncated). Solution: Run ffmpeg -list_devices true -f dshow -i dummy and copy the name verbatim, including spaces. Wrap in double quotes.

Error: x11grab ... BadMatch or black frame

Cause: $DISPLAY doesn't match, or running over SSH without -Y. Solution: export DISPLAY=:0.0 locally, or use ssh -Y and target the remote display.

Error: kmsgrab: Permission denied

Cause: kmsgrab requires CAPSYSADMIN. Solution: sudo ffmpeg ... or sudo setcap cap_sys_admin+ep $(which ffmpeg).

Output MP4 won't open / "moov atom not found"

Cause: ffmpeg was killed with SIGKILL / Ctrl+C mid-recording; MOOV never written. Solution: Always press q to stop. For future recordings use MKV, then ffmpeg -i in.mkv -c copy out.mp4. Try recovery: untrunc or ffmpeg -err_detect ignore_err -i broken.mp4 -c copy fixed.mp4.

Webcam stuck at 5fps (Linux)

Cause: v4l2 defaulted to raw YUYV; USB 2.0 bandwidth cap. Solution: Add -input_format mjpeg before -i /dev/video0.

Audio and video out of sync after a long recording

Cause: Clock drift between capture devices. Solution: Add -use_wallclock_as_timestamps 1 on each input, or record audio separately and mux after with -itsoffset to align.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: damionrashford
• Source: damionrashford/media-os
• License: MIT
• Homepage: https://damionrashford.github.io/media-os/

Install and usage instructions live in the source repository linked above.