1
Find a direction
Search scenes by role, technique, duration and style. Compare contrasting references instead of copying one.
MCP server · Private beta
Real launch references. Ready for your agent.
Your agent searches reviewed launch scenes, reads their timing, text, camera and sound, then builds something that belongs to your product.
- Claude
- Cursor
- ChatGPT
- >_Codex
Your agentconnecting…
0 MCP calls0 scenes read0.0 s
A real exchange with the library, replayed. Hover to pause.
18
tools
5
guided prompts
100%
reviewed scenes only
0
media files shared
Search like your agent does.
Describe a scene. Real results come back from the library, with the response your agent would read.
HookFramer1.2 s
HookFramer1.0 s
HookSuperhuman3.7 s
HookResend5.0 s
HookSuno2.4 s
HookTana1.8 s
Real results from the published library. The agent view is simplified; full responses add sources, verification and missing-data reasons.
From a brief to a film plan.
2
Read the measurements
Open a recipe: text timing, relative boxes, camera movement, audio and frames, each with its source and limits.
3
Make it your own
Plan a structure, adapt a recipe, compute a beat grid. Every extrapolation is labelled as one.
Guided prompts
/launchscene-film
Build a launch-film plan from measured, contrasting references.
/launchscene-hook
Explore five contrasting openings for a product and promise.
/launchscene-like
Adapt the structure and motion logic of a chosen reference.
/launchscene-review
Discuss a storyboard in the context of corpus variation.
/launchscene-vertical
Adapt a wide composition to a vertical canvas.
18 tools, four jobs.
Find, inspect, measure and keep. Pick a tool to see what it returns.
search_scenes
Compact scene cards from the site taxonomy and text search. Unknown measurements are excluded, never guessed.
Example call
search_scenes({ query: "typographic hook", filters: { role: "hook" }, limit: 6 })Returns
- Scene id, brand and video
- Role, start and duration
- Techniques, camera, on-screen text
- Source link and relevance score
Plug it in. 30 seconds.
Access is by invitation. Paste your private key into any client that supports Streamable HTTP.
Terminal
claude mcp add --transport http launchscene https://launchscene.app/api/mcp \
--header "Authorization: Bearer YOUR_PRIVATE_BETA_KEY"Streamable HTTP with a bearer header. Keep your key in local client settings, never in a shared repository.
Account-linked clients can save references to private boards shared with the website.
What’s measured, honestly.
Every reference carries its source, how it was checked, and what is missing. Your agent knows the limits too.
Text boxes are observations
A box is the position of the text at a recorded OCR frame. Older analyses may only give a midpoint estimate.
Easing is a label, not a curve
Camera moves are measured; the easing class is an estimate of their shape, never the original keyframes.
Reviewed is not every pixel
A reviewed video was checked by a human, scene by scene. Not every OCR box or beat was checked one by one.
Statistics describe, they don’t judge
Corpus numbers come with their sample size. They show how launches vary, not what is good.
Opening in waves
Give your agent real references.
Leave your email. You’ll get a private key as soon as your wave opens.