@attentionmachine/cli

The Attention Machine's video helper. It runs on your own computer and does the physical work of video editing: it measures a recording, uploads its audio, renders the finished cut, checks the result, adds captions, searches local overlays, and puts a finished video on a temporary link for posting. It never makes editing decisions — those happen on The Attention Machine's servers. Overlay search uses a small local model; frames, searches, camera places, and vectors never upload.

You don't install this yourself. Your AI runs it for you with npx when you ask for a video edit, using exact commands from The Attention Machine.

am doctor                                         is this computer ready?
am prepare "<clips…>" --edit ID --upload "URL"    measure + upload audio
                                                  (one video, several, or a folder)
am adopt --edit NEW --from OLD                    a new version reuses the recording
am render --edit ID --plan "URL" -o "OUT"         render the cut plan
am check --edit ID                                check the render, logged for us
am audit --edit ID --upload "URL"                 upload finished audio
am deliver --edit ID -o "OUT"                     move the finished video into place
                                                  (every video lands as .mp4, whatever -o says)
am captions prepare --edit ID [--upload "URL"]    cut sheet + measure for captions
am captions doctor                                ready to render captions?
am layers add --edit ID --file "F.html" --at S --for S
                                                  one layer file over a window — cards,
                                                  filters, footage, photos, zoom-ins are
                                                  all files in machine/layers/
am layers list|edit|remove|clear|confirm          see or change the layer plan
am layers --force                                 take newer stock files in machine/layers/
am layers index                                   remember what is in overlays/
am layers find "people on stage"                  find overlays by what the camera sees
am cuts --edit ID                                 every cut: trims to leave, retakes worth hiding
am sheet "<clip>" [--every S | --around S] [-o OUT.jpg]
                                                  one contact-sheet image of a local clip
                                                  (--around: one tile a second at that moment)
am deliver --edit ID --captions [--plan "URL"]    render every layer once + check sheet
am deliver --edit ID --captions -o "OUT"          move that rendered file into place
am deliver --edit ID --layers [-o "OUT"]          same, for a video with layers, no captions
am clean "<video>" --job ID --upload "URL"        clean the sound, rebuild the video
am publish --edit ID --upload "URL"               temporary public link for posting

Requires a Mac with an Apple chip (ffmpeg from Homebrew: brew install ffmpeg-full) or a Windows 10/11 computer (winget install ffmpeg --source winget --accept-source-agreements --accept-package-agreements), plus Node 20+. Rendering captions additionally needs Node 22+ (the renderer downloads a private browser on first run).

The first overlay search downloads a pinned local model (about 160 MB) into .attentionmachine/models/. Later searches reuse it. Frames, camera facts, vectors, and search words stay on the computer.