| name | omniget-fetch |
| description | This skill should be used when the user pastes a video, audio or social-post URL (YouTube, TikTok, Instagram, X/Twitter, Reddit, Vimeo, Twitch, Bilibili, Threads, Pinterest, direct .mp4/.m3u8 links) and asks to "download this", "get this video", "save the audio", "grab this reel", "fetch this clip", "pull the mp4", or otherwise wants the media file on disk. Also covers "what tools do I have for downloading" and "why did the download fail". |
| version | 0.1.0 |
Fetching media with OmniGet's tooling
Download media through the plugin's scripts instead of hand-writing yt-dlp
invocations. The scripts reuse whatever the user already has, in this order:
omniget-cli on PATH, the binaries OmniGet manages inside its app-data folder,
then system yt-dlp/ffmpeg. Every script prints one JSON line on success.
Scripts live in ${CLAUDE_PLUGIN_ROOT}/scripts/.
Workflow
- Run the download. Quote the URL.
bash "${CLAUDE_PLUGIN_ROOT}/scripts/fetch.sh" "<url>" [--audio] [--quality 720] [--out DIR]
Output: {"file","title","duration","size","platform","engine","id"}.
- Default output directory is
~/Downloads/omniget (override with OMNIGET_DIR or --out).
--audio keeps only the audio track as .m4a. --quality N caps the video height.- Instagram, Threads, X and Bilibili links go through
omniget-cli when it is installed,
which brings OmniGet's native extractors and cookie accounts.
- Report the file path, size and duration in one or two sentences. Do not paste the JSON.
- If the user only wants to know what the media says, use the
omniget-transcribe skill
instead of downloading video.
First run / missing tools
If fetch.sh reports yt-dlp not found (or a fresh machine has nothing set up),
do not hand the user raw install commands. Offer one step:
bash "${CLAUDE_PLUGIN_ROOT}/scripts/setup.sh"
It detects the OS, lists what is missing, and installs it after a single confirmation
(Mac via brew, Windows via winget/scoop; Linux runs the distro manager when it can, else
prints the commands). Pass --yes to skip the prompt. If OmniGet (the desktop app) is
installed, yt-dlp and ffmpeg are already managed and nothing needs installing.
When something fails
Run the doctor first, then act on what it reports:
bash "${CLAUDE_PLUGIN_ROOT}/scripts/doctor.sh" --check-keys
| Symptom in stderr | Meaning | Action |
|---|
yt-dlp not found | no engine available | show the doctor's install lines and ask before installing anything |
Sign in to confirm, login required, Private video, HTTP 401/403 | platform wants a session | use OmniGet's cookies (automatic when the app has an account for that site) or set OMNIGET_COOKIES_FROM_BROWSER=chrome and retry |
HTTP 429, rate-limit | too many requests | wait 30 s and retry once; then stop and tell the user |
empty media response, HTTP 400 (Instagram/X) | rate-limited or login-gated | wait a few minutes and retry; if private, provide cookies (see below). omniget-cli (installed by setup) handles these sites better than raw yt-dlp |
DRM, SAMPLE-AES, Widevine | protected stream | stop; explain that OmniGet does not bypass DRM |
Unsupported URL | no extractor | say so; suggest the page's direct media link if the user has one |
Requested format is not available | quality cap too strict | retry without --quality |
On failure the scripts print a plain-language -> hint line classifying the cause (rate-limit / login / DRM). If a site that used to work now fails for everyone, the fix is usually stale tools: bash "${CLAUDE_PLUGIN_ROOT}/scripts/setup.sh" --update.
Automatic login retry. When a fetch is blocked by a login wall or rate-limit, the scripts retry once on their own using the browser you're logged into (auto-detected: Chrome/Brave/Edge/Firefox/Safari). No flag needed. Two things to know: on macOS the first browser-cookie read raises a one-time Keychain prompt (allow it, or choose Always Allow); and if your browser has several profiles and the default one isn't logged in, pin the right one — OMNIGET_COOKIES_FROM_BROWSER=chrome:"Profile 3". If it's still blocked after the retry, it's a hard rate-limit — wait a few minutes.
Rules
- Downloads happen only when the user asked for the media. A bare URL with a question
about its content is a transcription request, not a download request.
- Never run an install command (brew, winget, apt, pip, uv) without the user's explicit yes;
the doctor prints the commands, the user decides.
- Cookies files and API keys are never printed, copied or committed.
- Prefer
--audio for podcasts, talks and anything the user will only listen to or transcribe.
Environment variables
OMNIGET_DIR (output), OMNIGET_DATA_DIR (OmniGet app data, auto-detected per OS),
OMNIGET_TOOL_YT_DLP / OMNIGET_TOOL_FFMPEG (explicit binaries),
OMNIGET_COOKIES_FROM_BROWSER (chrome, firefox, safari, edge).
