Find and download licensed media from 9 free sources

web-media (webmedia.py) queries multiple free image, video, and GIF sources in a single fan-out and returns a normalized result list, with an optional download of the top-K results plus an attribution sidecar (attribution.json: source, author, license, url, page_url). It is a zero-dependency stdlib Python script. Nine sources are supported: five need no API key - openverse (CC web images from Flickr and museums), wikimedia (factual, historical, or landmark photos), internetarchive (historical and archival images and films), loc (historical US prints and photos), and nasa (space imagery and video) - and four need a free key - pexels and pixabay for modern stock photos and short video clips, and klipy or giphy for GIFs. klipy is the recommended GIF source (free, unlimited, a Tenor drop-in), though its adapter is unverified and assumes Tenor-compatible request/response, worth checking against klipy's own docs once keyed; the previous tenor adapter was removed after Google ended that API on 2026-06-30. Each result follows a fixed record schema: {source, title, url, thumb, dl, page_url, author, license, w, h, type}, where dl is the directly downloadable media URL and is None when only a web page exists.

Use it when a task needs a real or archival photo or clip - a hero image, a texture, a visual reference, historical footage, or a reaction or animated GIF - rather than a generated one. Keep the video caveat in mind: archival sources like Internet Archive or the Library of Congress host whole films or documentaries, not single shots, so a modern single clip should come from pexels or pixabay instead, since those are born as short clips with a direct MP4. Getting a single historical shot out of an archival film needs an extra step: either Twelve Labs' Marengo search (a free 600-minute tier) against the film's public MP4 URL to get a timestamped moment, clipped with ffmpeg, or PySceneDetect to cut the film into shots locally and rank keyframes with CLIP via the sibling muser skill. GIF sources only fire with --type gif. This skill is the internet-retrieval capability specifically - a peer to muser (local semantic search) and fal (generation), not a replacement for either; a future router fanning out across all three isn't worth building until the model demonstrably mis-routes without one.

Input is a text query plus flags: --source all (default), nokey (no-key sources only), or a comma list like wikimedia,pexels; --type image|video|gif; --count N; --json; and --download --out DIR to fetch each result's direct media URL alongside the attribution sidecar. A documented usage example:

webmedia.py "1950s street scene" --type image --count 8 --json
webmedia.py "rocket launch" --type video --source nasa,internetarchive
webmedia.py "car factory 1930s" --source all --download --out /tmp/cars

For sound effects rather than images or video, two sibling scripts in central/scripts/ cover audio: freesound-fetch.py searches freesound.org and downloads short hq-mp3 previews with license and user metadata for attribution, and audio-judge.py sends a clip to OpenAI's audio-native gpt-audio model and returns {heard, score, matches, suggestion} for a generate/fetch-judge-iterate loop - useful for culling obvious mismatches, but not a trustworthy judge of subjective qualities like "grating" (it tends to label nearly any beep "sharp/high-pitched"), so a final aesthetic call should still be confirmed by ear.

Optional keys go in central/.env: PEXELS_API_KEY and PIXABAY_API_KEY for their respective stock sources, KLIPY_API_KEY and GIPHY_API_KEY for GIFs, and FREESOUND_API_KEY for the sibling audio-fetch script; audio-judge.py auto-sources a real sk--prefixed OPENAI_API_KEY from the same .env and deliberately ignores a local lm-studio stub env var. It sits alongside two peer skills in the same toolkit, muser (local semantic search) and fal (generation).

Anyone building on a coding agent who needs a real or historical photo, video clip, or reaction GIF pulled from the open web with license and attribution intact, rather than generating one from scratch - including retrieving a specific shot out of an archival film, or fetching and machine-screening a sound effect before final human review. It is MIT licensed.