ScreenBeaver/ScreenBeaver

Screen Beaver live demo builder

The user wants to sit idle while you drive their product (smooth cursor + clicks + real typing) and build a Screen Beaver video. Prefer native capture (no Windows share picker). Edits apply in auto mode (one recipe). They click Export themselves.

Hard rules

1. Use screen-beaver MCP only. Prefer record_and_build (includes demo automation). Use atomic demo_* tools or run_demo for custom click paths.
2. If screen-beaver is missing from this session’s MCP tool list, stop. Tell the user to enable it under Customize → MCPs, run npm run mcp:install once, then reload Cursor / start a new chat. Do not fall back to raw HTTP, curl, fetch, or Shell against 127.0.0.1:17355.
3. Never write or execute .ps1, .bat, .cmd, or other shell scripts to talk to Screen Beaver.
4. Never create JSON demo/recipe files on disk as a workaround — pass steps/payloads in MCP tool arguments.
5. Do not call export_mp4 unless the user explicitly says to export.
6. Keep Screen Beaver desktop open (npm run desktop:dev or the installed app). Bridge: http://127.0.0.1:17355 (used by the MCP server only). If ping_studio fails, tell them to open the app — do not hang.
7. Never ask them to click the OS screen-share picker — use start_capture / record_and_build (native window select).
8. Never idle-record a blank page and only sprinkle captions. Always drive the UI during capture.
9. Keep takes as short as the script needs. Stop as soon as the last click/type finishes (durationSec pad ≤ ~0.4s). Do not add long trailing waits.
10. Use demo_type / { op: "type", text } for typing — key-by-key, never clipboard paste.
11. Narrate briefly in chat so they watch the Screen Beaver window.
12. Coords: call get_window_layout (or use contentNx/contentNy) — never trust Cursor’s embedded browser size. Page-relative contentNx/contentNy map through browser chrome automatically.
13. Auto mode: let the studio derive zooms + typing SFX from real clicks/keystrokes. Do not hand-place agent zooms/sounds that fight Auto. record_and_build captions only + regenerateAutoZooms.
14. Voiceover: after captions, call generate_voiceover (default English). If the user asks for another language, pass targetLang (e.g. hi-IN) — captions are translated first, then Edge TTS. Prefer record_and_build({ voiceover: true, voiceoverLang }) which does this automatically.
15. Background music: use Openverse via MCP (pick_background_music / record_and_build with backgroundMusic: true). No Pixabay / .env keys. Glare / clear effects play the bundled shimmer SFX (Shining.mp3).

Workflow A — full hands-off (opens Screen Beaver UI first)

record_and_build now:

1. focus_studio — brings Screen Beaver to front
2. open_record_picker — opens the same “Start a new recording” / share options UI
3. Opens productUrl / app
4. start_capture — selects the window natively (Application), shows Recording…
5. Runs demoSteps
6. Stops immediately + auto zooms/captions from the take
7. generate_voiceover — English by default; translate+TTS when voiceoverLang is set
8. pick_background_music — Openverse CC bed (soft corporate instrumental)

ping_studio
  → record_and_build({
      productName, productUrl, titleContains: "Edge",
      voiceover: true,
      voiceoverLang: "en-US",   // or "hi-IN" / "es-ES" etc.
      backgroundMusic: true,
      musicQuery: "upbeat corporate instrumental",
      demoSteps: [ ... ],
      introCaption: "...",
      features: [{ title: "Sign in", atSeconds: 3 }]
    })
  → Tell user: click Export when ready

Or on an existing take: generate_voiceover({ targetLang: "en-US" }) pick_background_music({ query: "calm tech ambient instrumental" }) generate_voiceover({ speakLang: "en-US", targetLang: "hi-IN" })

Demo steps

op	meaning
wait	{ ms } — keep short; no long end pads
move / click	prefer { contentNx, contentNy } 0..1 in webpage; or { nx, ny } outer window
type	{ text, delayMs? } key-by-key (emits typing SFX)
scroll / focus / right_click	as before

Limits

• Native capture + input automation are Windows desktop only.
• Demo mouse/keyboard input only works while capture is running.
• The OS Share dialog is never shown to agents — Screen Beaver flashes the in-app share options, then captures the window natively.