From Cursor chat to polished MP4: MCP product demos in practice

The appealing version of agent-made demos sounds like one command: ask Cursor for a walkthrough and receive a finished MP4. The practical version is better, because you can see and control what happens between those points. Screen Beaver’s MCP package gives the agent recording and studio actions, while the desktop app remains the place where you review the take. `record_and_build` handles a defined flow in one run. `live_build_demo` is useful when you want to watch the tools fire and steer the process step by step.

Start with a boring connection test

Install `@screenbeaver/mcp`, add the server command to your MCP client, open Screen Beaver, and test against a harmless page. Ask the agent to start a short capture, click one known control, stop, and leave the project open. This catches configuration errors before a ten-step recording. It also confirms which display or window is being captured, whether Windows scaling keeps the type sharp, and whether menus remain inside the frame. Review what the agent could access before widening the task. Don’t make your first run the production billing flow with real customer names visible; that is an avoidable way to learn how permissions work.

Choose the right build mode

Use `record_and_build` when the path is stable and you can describe it precisely: open the seeded account, create an alert, trigger it, and show the result. It can capture the actions and apply the first polish pass without requiring commentary after every click. Use `live_build_demo` when the route is exploratory, when a modal may vary, or when you want to approve each scene. The live option takes more attention, but it’s easier to correct early than to discover at the end that the agent demonstrated the wrong plan.

• State the exact URL, account, and starting screen
• Describe visible outcomes instead of vague goals
• Tell the agent which fields contain safe sample data
• Request pauses after navigation, saves, and generated results
• Set a clear stopping point so the take doesn’t wander

Prompt for a visual story

An agent can complete a task correctly and still make a terrible demo. If you ask it to “show team permissions,” it may jump straight to a settings URL, toggle a role, and stop. A viewer needs context. Ask it to begin on the team page, open one member, change the role, then return to the list so the updated badge is visible. Include pacing notes, the intended audience, and the question each scene answers. A developer tutorial can tolerate more setup; a homepage clip should reach the payoff almost immediately. Tell the agent to pause on evidence rather than after every click, because pauses are editorial beats, not generic delays.

Inspect what the automation produced

Screen Beaver leaves the result editable. Watch the pointer path and verify every typed value. Then check the generated zooms: did the frame emphasize the decision, or merely the largest click target? Regenerate captions if the first copy describes mechanics instead of value. Add Edge TTS only if narration improves the channel you’re publishing to. I usually keep the first automated take as a diagnostic, tighten the prompt, and record once more. The second run is faster than rescuing a fundamentally confusing sequence.

Plan for failures without fearing them

Agents can misread duplicate buttons, arrive before content loads, or encounter a staging banner that wasn’t there yesterday. That’s not a reason to avoid MCP; it’s a reason to use deterministic fixtures and short scenes. Seed the account, stabilize the viewport, and split a long tutorial into sections that can be rerun independently. If authentication requires a passkey or CAPTCHA, handle that boundary manually and let the agent continue afterward. Some steps should remain human.

Finish and export on Windows

Once the cut feels right, trim dead time, settle caption timing, and export H.264 locally. Watch the exported file rather than assuming timeline playback is enough; compression, audio balance, and caption edges can look different in the actual asset. The file does not need a cloud encode, and the project remains available for the next release. That ownership is important: MCP makes repetition cheap, but your editorial choices are still the valuable part. Save the final prompt, sample-data reset, and capture settings with the project so another teammate can rerun the walkthrough later. The useful system is not one impressive agent trick; it is a repeatable path from product change to reviewed MP4.