Watch a local specialty coffee roaster talk passionately about anaerobic Colombian beans while an auto-editing web app cuts away to stock footage of two strangers in suits high-fiving in front of a glass skyscraper.
Viewers swipe away in half a second because generic keyword stock clips smell like a dropshipping ad from 2021.
Real businesses already sit inside rooms full of great B-roll. A Pilates studio has reformers sliding, chalked hands, and morning sunlight hitting oak floors; a contractor has tile joints, laser levels, and freshly grouted showers. What kills consistency is not filming three-second clips on a phone, it is sitting in a timeline editor figuring out which clip belongs on second 0:06 versus second 0:19.
Mixbash’s plan_visuals stage on https://mixi.mixbash.com/api/agent/mcp turns B-roll matching into a structured producer workflow that works in two directions.
In the footage-first path, you normalize a handful of raw phone clips locally with mixbash-intake-v1 (1080p H.264, CRF 23, AAC), upload them via prepare_upload and complete_upload, and call register_media with concise shot descriptions alongside your talking-head clip’s Whisper word timestamps.
When your agent calls plan_visuals, Mixbash aligns your actual uploaded B-roll assets against exact spoken phrases, placing the espresso pour right when you mention extraction temperature and holding on your face during the punchline.
In the script-first path, you run create_project and draft_script before filming anything at all, then call plan_visuals to generate a concrete, beat-by-beat Shot List. Hand that five-bullet checklist to a junior staff member on-site, and they can capture every required three-second phone clip in under eight minutes without guessing.
Because plan_visuals deals in clean JSON beat records rather than streaming raw video frames into your chat prompt, running the whole planning loop on GPT-6.2 Sol ($2/$10), Claude Sonnet 5.5 ($2/$10), or Haiku 5.5 ($0.10/$0.50) costs a fraction of a penny.
Here is the honest catch: plan_visuals will not fabricate fake AI-generated video of products you do not sell just to fill a gap. Someone on your team still has to point a phone camera at your actual shop, clinic, or software screen for a few minutes each week, which is precisely why the finished Reel looks unmistakably like your business.