Agent Loop Short Video Script
DeskVNC: 90-second video script for the four-call loop
For: Maintainer to record Format: 90 seconds total Channel: YouTube
Shot List
| Shot | Duration | Type |
|---|---|---|
| 1 | 0:00 to 0:05 | Full-screen terminal, DeskVNC MCP registration + dvv_hosts output |
| 2 | 0:05 to 0:10 | Terminal: dvv_open with perceive=true, remote desktop window appears |
| 3 | 0:10 to 0:14 | Terminal: dvv_control acquire, pointer on remote desktop becomes agent cursor |
| 4 | 0:14 to 0:18 | Terminal: dvv_screen at scale 0.25, thumbnail of the remote desktop |
| 5 | 0:18 to 0:25 | Terminal: dvv_click, notepad opens on remote desktop; dvv_type, text appears |
| 6 | 0:25 to 0:32 | Terminal: second loop cycle, different UI element; dvv_screen damage-crop |
| 7 | 0:32 to 0:40 | Human moves the mouse; LEASE_REVOKED printed; human cursor is shown |
| 8 | 0:40 to 0:45 | Terminal: dvv_close; remote desktop pane disappears |
| 9 | 0:45 to 0:52 | Performance overlay: 19 ms observe-then-act, 447 chars per second typing, 4 ms open |
| 10 | 0:52 to 0:58 | Code panel: the four calls in sequence, with tool names highlighted |
| 11 | 0:58 to 1:02 | Boundary: attended support option shown as a parallel use case |
| 12 | 1:02 to 1:08 | Product shots: VNC, RDP, SSH all running in the same DeskVNC window |
| 13 | 1:08 to 1:22 | GitHub page and install link; CTA to try it today |
| 14 | 1:22 to 1:30 | Outro card: github.com/psmux/DeskVNC |
Narration and On-Screen Text
Shot 1 (0:00 to 0:05)
On screen: Dark terminal on left; DeskVNC window on right showing a saved host library with live thumbnails.
On-screen text (bottom third):
DeskVNC MCP Server: dvvNarration:
Most tools that put an AI agent on a remote desktop require an agent running on that desktop. That is refused on Citrix, on VDI, on jump hosts, and on anything a client owns. DeskVNC speaks the protocol those machines already speak. Nothing new on the far end.
Shot 2 (0:05 to 0:10)
On screen: Terminal window, large. Agent dashboard on the right.
On-screen text (lower third, monospace, syntax highlighted):
dvv_hosts {}
-> [ { hostId: "win-dev-01", name: "Dev Desktop", protocol: "RDP" } ]Narration:
Step one. Ask what is there. dvv_hosts returns the saved machines from the library.
Shot 3 (0:10 to 0:15)
On screen: Terminal with the open call.
On-screen text (lower third):
dvv_open { "hostId": "win-dev-01", "perceive": true }
-> { limbId: "dvv://win-dev-01/desktop", width: 1920, height: 1080 }Narration:
Step two. Open the machine with perceive set to true. The remote desktop appears in DeskVNC. The call returns a limb id and the screen dimensions, 1920 by 1080.
Shot 4 (0:15 to 0:20)
On screen: Terminal with control acquire.
On-screen text (lower third):
dvv_control { "limbId": "dvv://win-dev-01/desktop", "action": "acquire" }
-> { lease: { holder: "agent", acquired: "2026-10-08T..." } }Narration:
Step three. Take the wheel. dvv_control acquire is under 1 millisecond. The agent now holds the lease on this limb.
Shot 5 (0:20 to 0:28)
On screen: Terminal with screen read, then thumbnail of the remote desktop in the agent panel.
On-screen text (lower third):
dvv_screen { "limbId": "...", "form": "full", "scale": 0.25 }
-> { imageSpace: { x:0, y:0, w:480, h:270, scale:0.25 }, generation: 7 }Narration:
Step four. Read the screen at scale 0.25. This is 25 milliseconds and returns a 112 kilobyte image. The generation number is printed on every screen read. That number fences every click and every keystroke that follows.
Shot 6 (0:28 to 0:38)
On screen: Three calls in rapid sequence, each with a cut to the remote desktop reacting.
On-screen text (lower third, three panels cycling):
Panel 1:
dvv_click { "limbId": "...", "x": 700, "y": 400, "generation": 7 }
-> { delivered: true }
Panel 2:
dvv_type { "limbId": "...", "text": "hello world", "wpm": 3000 }
-> { delivered: 11, elapsed_ms: 25 }
Panel 3:
dvv_key { "limbId": "...", "keys": "meta+r" }
-> { delivered: true }Narration:
Click at (700, 400). Type 11 characters at 3000 words per minute. Press meta-r to open the run dialog. Each click carries the generation from the last screen read. A click against a stale generation is refused rather than landing somewhere unintended. One observe-then-act cycle takes 19 milliseconds. That is about 52 actions per second.
Shot 7 (0:38 to 0:48)
On screen: Remote desktop. Someone moves the mouse. The terminal shows LEASE_REVOKED.
On-screen text (lower third, red accent):
dvv_status { "limbId": "..." }
-> { lease: { holder: "human", human_took_over: true } }
SCREEN_CHANGED: a person moved the mouseNarration:
A person can take the wheel back at any moment. A human move prints LEASE_REVOKED and sets human_took_over to true. The right move for the agent is to stop. Held keys and buttons are released, so a half-finished drag never strands the desktop.
Shot 8 (0:48 to 0:54)
On screen: Terminal with close call.
On-screen text (lower third):
dvv_close { "limbId": "dvv://win-dev-01/desktop" }
-> { state: "closed" }Narration:
When the task is done, close the limb. The remote desktop pane closes. The lease is released.
Shot 9 (0:54 to 1:02)
On screen: Split panel. Left: DeskVNC running three simultaneous machines, VNC, RDP and SSH. Right: the full four-call sequence in JSONC.
On-screen text (left panel label):
VNC RDP SSHOn-screen text (right panel):
dvv_hosts {}
dvv_open { "hostId": "...", "perceive": true }
dvv_control { "limbId": "...", "action": "acquire" }
dvv_screen { "limbId": "...", "form": "full", "scale": 0.25 }Narration:
VNC, RDP and SSH all in the same client. Each machine is its own limb with its own lease, so ten machines are ten independent loops running in parallel.
Shot 10 (1:02 to 1:10)
On screen: Performance benchmark overlay.
On-screen text (benchmark card, centered):
4 ms dvv_open and attach
<1 ms dvv_control acquire
25 ms dvv_screen at scale 0.25
19 ms one observe-then-act cycle
447 char/s dvv_type throughputNarration:
Benchmarks from a real 1920 by 1080 Windows desktop over LAN. dvv_open and attach in 4 milliseconds. Acquire in under 1 millisecond. A screen read at scale 0.25 in 25 milliseconds. And typing at 447 characters per second.
Shot 11 (1:10 to 1:18)
On screen: GitHub page, release download section.
On-screen text (footer bar):
github.com/psmux/DeskVNC
github.com/psmux/DeskVNC/releases/latestNarration:
DeskVNC is free and open source, MIT or Apache-2.0. Download the latest release, register the dvv MCP server, and an agent can be driving a real desktop within minutes.
Shot 12 (1:18 to 1:30)
On screen: Full-width outro card. Dark background.
On-screen text (large, centered):
github.com/psmux/DeskVNC
VNC. RDP. SSH.
Nothing installed on the far end.
A person can take the wheel back at any moment.Narration (quiet, fade out):
One native client for every machine you look after.
Technical Notes for Recording
- Record at 1920 by 1080, 30 fps minimum.
- Use a real DeskVNC session for the demo footage. Screen record the terminal running real dvv commands and the remote desktop reacting in real time.
- If the remote desktop is not available, use a local VM on the same LAN to keep the latency realistic.
- The benchmark figures in Shot 10 are verified from the project README. Do not round, extrapolate or change them.
- Keep all four real error codes present in the narration: LEASE_REVOKED, GEOMETRY_CHANGED, SCREEN_CHANGED, LIMB_GONE. No other refusal code name is in the protocol, so do not invent any.
- Do not narrate any code name that is not one of the four above.
- Positive framing only. Do not mention what the product lacks.