AIOProductOS Studio · MCP
Your AI teammate becomes your product videographer. Studio is an MCP server
that records scripted walkthroughs of your own web app — a visible cursor
that glides to every click, camera punch-ins, highlight callouts, storyline
captions, designed scene transitions — and takes marketing-grade screenshots.
Tell your AI host what story to shoot; get back a share-ready MP4.
The videos PMs hate making — demo clips, explainers, changelog GIFs, launch
screenshots — become one prompt.
Free · MIT · runs 100% locally (your browser, your URLs, your logged-in
session — nothing is uploaded anywhere). By AIOProductOS,
built from the same pipeline we use for our own launch videos.
Install
claude mcp add studio -- npx -y @aioproductoscom/mcp-studio
{ "mcpServers": { "studio": { "command": "npx", "args": ["-y", "@aioproductoscom/mcp-studio"] } } }
Requirements: Node 18+, Playwright Chromium (npx playwright install chromium,
one-time), and ffmpeg on PATH for MP4/GIF export (brew install ffmpeg — without
it you still get the raw WebM).
Use
Then just direct the film:
"Record a 30-second walkthrough of https://app.example.com — theme it to our
brand (#0d1117 bg, #f97316 accent), open on the dashboard with the title
'Meet Example', caption the three key features, zoom in on the analytics
chart, and end with 'Start free · example.com'. Also grab a hero screenshot."
Your host runs the shoot tool by tool; files land in ./studio-output/.
The tools
studio_start | open a themed recording browser (viewport, brand colors, logged-in state) |
studio_goto | navigate behind a designed title card — no loading flashes on tape |
studio_click / studio_type | cursor glides to the target; typing is human-paced |
studio_scroll | cubic-eased scroll, never a jump cut |
studio_caption | bottom-left storyline caption — also the spoken voice-over script |
studio_narrate | record a spoken line with no on-screen caption (narration over action) |
studio_highlight | accent ring + label callout around any element or visible text |
studio_zoom | smooth camera punch-in on a metric, button, or result |
studio_hold | hold the shot (micro-drift keeps frames flowing) |
studio_end_card | closing glass card: title, subtitle, CTA URL |
studio_screenshot | high-DPI PNG of the frame, an element, or the full page |
studio_finish | stop, auto-remove dark frames, export MP4 — plus voice-over, subtitles, music, social crops, GIF |
studio_cancel | abandon the take |
Sound & social (v0.2)
Your captions are the script. studio_finish turns them into a finished, shareable cut:
studio_finish({
name: "onboarding",
voiceover: true, // narrate the caption lines with a LOCAL voice — no cloud, no API key
voice: "Samantha", // any installed system voice (optional)
subtitles: "srt", // ".srt" sidecar from the captions · "burn" to render them in-frame
music: "./bed.mp3", // optional local music bed, ducked under the narration
aspects: ["9:16", "1:1"], // extra Reels/TikTok/feed crops beside the 16:9 master
gif: true,
})
Voice-over is 100% local and offline — it uses macOS say, piper
(set STUDIO_PIPER_MODEL to a downloaded .onnx voice), or espeak-ng, whichever it finds. No account,
no key, nothing uploaded. No engine on the machine → the video is simply silent (with a hint how to enable it).
Narration timing is auto-corrected for the frames the dark-frame trimmer removes, so speech stays in sync.
Recording logged-in areas
Studio never sees your credentials. Save a Playwright storageState
once (npx playwright codegen --save-storage=auth.json https://app.example.com),
then pass storage_state_path: "auth.json" to studio_start.
Keeping junk off the tape
- Cookie banner? Pass your consent cookie via
cookies so it never mounts.
- Chat bubble / dev overlay? List selectors in
hide_selectors — hidden from frame 1.
- Wrong flash color on page loads? Set
theme.bg to your app's real background.
Why it looks good (the hard-won part)
No white flash on page commits (the theme background paints before any page CSS),
no black frames (navigations hide behind designed title cards, and residual dark
frames are detected and cut deterministically with ffmpeg), no stutter (every
camera move is compositor-friendly easing), and action that reads on camera
(a visible cursor glides to every target before it acts).
Links
MIT © AIOProductOS Inc.