Generate a launch video
🔒Try a prompt-injection idle
Three demonstrations of the agent refusing hostile input. Each runs the same end-to-end loop and surfaces the guardrail notes.
Live agent output
0 events
Storyboard
Agent loop trace
⚙️Agent config (AdaL Cloud)
Name
…loading…
Model
…loading…
DRY_RUN
…loading…
Allowed origin (CORS)
…loading…
Render profile
…loading…
Positive tool allowlist
Forbidden tools (by absence)
Bash · Edit · Read · Search · Web · Image · Video · Consult
System prompt (excerpt)
🛡️Guardrails
🏗️Architecture
1
User enters URL + brief in this Cloudflare Pages UI.
2
BFF creates an AdaL Cloud session + forwards the chat (SSE).
3
Agent calls its 6 scoped tools (no Bash, no Web).
4
Composition HTML is validated and rendered by HyperFrames.
5
Rendered MP4 streams back via
/output/<job>.mp4.
Cloudflare Pages hosts only the static UI. The BFF + renderer run on a normal Node+FFmpeg container (Render.com or AWS App Runner) because Cloudflare Workers cannot run FFmpeg or headless Chrome.