Application demonstrations

Screenshots come from the running Grokbot Harness.app, with actual Agent replies. Capture uses an isolated Docker Box, data directory, and desktop profile, without existing user conversations or credentials.

File	Contents
workspace.png	Isolated local workspace
compose.png	Task entered through the desktop
agent-result.png	Reply after real-model execution
agent-demo.gif	Time-compressed README animation
agent-demo.mp4	The same sampled frames as H.264 video
capture.json	Versions, model, task, sampling, verified file output

Task

Create a launch checklist for a local agent project, save it to /workspace/launch-checklist.md, read it back, and return a summary. The configured Responses API supplies inference; shell commands run in the Linux Box. The capture script requires both a real output file and an Agent reply before producing successful demonstration media.

Timing

Capture samples about one frame per second, adds a final hold, then encodes four source frames per second. Waiting time is compressed; GIF/MP4 duration is not a performance benchmark. No UI content is generated or replaced to simulate functionality.

Record again

Build Host and the .app, start the default local stack, and run from a terminal with a configured model key:

node runtime/tools/capture-showcase.cjs

Requires macOS, Docker, ffmpeg, a working model API, and a free local CDP port 19224. The script replaces demonstration media here; raw frames and logs stay under ignored .runtime/showcase. It cleans up only its temporary container and app process. Inspect images, video, and capture.json together before publishing.