OpenAI · Chooses, studies and plans
Plan a product launch video with ChatGPT
Add Kuchi in ChatGPT's Plugins and ChatGPT can search a library of real product-launch videos, show you the posters, and explain why each one suits your launch.
Once you choose, it reads the shot breakdown and writes a plan timed to the reference's frames, with your product, copy and assets in each shot. A coding agent such as Codex builds the video from that plan.
Step by step: How to add a custom MCP server to ChatGPT
Connect Kuchi to ChatGPT
No sign-in, no API key. The server is https://kuchi.kalarislabs.com/mcp.
In ChatGPT, open Plugins and select + → Add custom MCP server.
Open PluginsName it Kuchi, paste this URL as the Server URL and choose No authentication. Accept the warning and select Create as a plugin.
https://kuchi.kalarislabs.com/mcpStart a chat and paste the setup prompt from this page.
Why use ChatGPT for your launch video
- Quick to try: no terminal, nothing to install on your machine.
- Good at comparing templates and drafting the script and on-screen copy.
- Hands a plan to Codex, which can build and render it.
How it works
- 1
Choose a template
A short list of real launch videos ranked for your product, with posters and links.
kuchi_suggest_templates - 2
Study it
Every shot's exact frames, how it enters, how much moves, its colours and a sheet of real frames.
kuchi_get_breakdown - 3
Plan the remake
Your product details mapped onto the reference's frames, with a list of anything still missing.
kuchi_plan_remake - 4
Build
Hand the blueprint to a coding agent such as Claude Code or Codex, which builds and renders it.
your coding agent - 5
Check
Compare your frames with the reference at the times it returns, and fix what differs.
kuchi_check_remake
Prompts to try in ChatGPT
Use Kuchi to find launch videos for a consumer mobile app with phone mockups. Show me the posters.
Compare the top three: pacing, length and what I would need to supply.
Write the shot list for my app to the timings of the one I pick.
How ChatGPT loads the MCP server
ChatGPT adds remote MCP servers in Plugins: + → Add custom MCP server. You give a name, the Server URL and the authentication type; this server uses No authentication. ChatGPT warns that custom servers are not reviewed; accept and create it as a plugin.
ChatGPT connects over Streamable HTTP. Some ChatGPT surfaces only call two tools, search and fetch; this server provides both alongside the richer kuchi_ tools, so it works in either case.
On business workspaces an admin may need to allow custom MCP servers before members can add them.
A full run in ChatGPT, step by step
In a chat with the plugin on, ChatGPT calls kuchi_suggest_templates and shows the posters with links. After your pick it reads the breakdown, analyses the storyboard and calls kuchi_plan_remake. If you ask for a different length or aspect ratio, it passes target_seconds and aspect_ratio and the blueprint is re-timed on whole frames.
The blueprint is the deliverable: a frame-timed shot list with your content in each shot. Ask ChatGPT for it as JSON, then hand it to Codex or Claude Code to build and render.
Tool results come back as Markdown by default; ask for JSON when you want the structured version.
What ChatGPT sends to the server
The server is a remote MCP server over Streamable HTTP at https://kuchi.kalarislabs.com/mcp. Every message is a JSON-RPC 2.0 request sent as an HTTP POST to that one endpoint. It is stateless: there is no session to keep alive, no cookie and no Authorization header, so ChatGPT can reconnect at any time without losing anything.
When ChatGPT connects it sends initialize and receives the server's name, version, instructions and capabilities (tools, resources and argument completion), then calls tools/list. The tool list carries a one-hour cache hint, because it only changes when the server is deployed. Each tool call is a tools/call whose arguments are checked against a strict schema: unknown fields are rejected and every string has a length limit, so a malformed call fails with a clear message instead of a half-right answer.
Results come back as text, Markdown by default or JSON when the call passes response_format: "json", and JSON results also arrive as structuredContent. Media is never inlined: the reference MP4, the poster and the storyboard sheets come as resource_link entries pointing at the CDN, which ChatGPT can open or fetch itself. A result is capped at 25,000 characters; a long blueprint arrives in pages of up to 40 shots, with shot_offset and shot_limit to fetch the next page.
Every tool is annotated readOnlyHint: true, destructiveHint: false, idempotentHint: true and openWorldHint: false: nothing changes on the server, the same call returns the same answer, and no tool reaches out to the wider web. The server also offers resources (kuchi://workflow, kuchi://videos/{id}, kuchi://breakdowns/{id} and kuchi://components/{name}) with autocompletion, for clients that let you attach a resource to a message.
The 13 tools, and when ChatGPT calls them
| Tool | What it returns | When ChatGPT uses it |
|---|---|---|
kuchi_get_workflow | The remake method, whole or one step (find, study, plan, build, check) with the rules that hold for all steps. Markdown. | Read once at the start of the chat so the app follows the method. |
kuchi_suggest_templates | A short numbered list of launch videos ranked for a product description, each with poster, link, what fits, what differs, what you would need to supply and one recommendation. | First call of a job; the app shows the posters inline and waits for your pick. |
kuchi_search_videos | Searches the whole library by keyword, style, category, format, aspect ratio, length, pace or theme, a page at a time. | When you want to browse beyond the suggested few. |
kuchi_get_video | One video by id, slug, site link or original X link: file, poster, storyboard, length and measured rhythm, with resource links to the MP4 and images. | When you paste a link from the site. |
kuchi_list_filters | Every valid style, category, format, aspect ratio, pace and theme with counts, plus the 14 shot kinds an agent classifies shots into. | Before a filtered search. |
kuchi_get_breakdown | The frame-by-frame measurement of one video: each shot's start and end, entrance, amount of motion and colour roles, a per-second motion and loudness curve, and storyboard sheets of real frames. | Study step; the app reads the storyboard images and describes each shot. |
kuchi_plan_remake | Takes your product details, assets and the agent's shot analysis; returns a blueprint re-timed to your length on whole frames, with your content per shot, plus missing_inputs and ready. | Called until ready=true; the blueprint is what you hand to a coding agent. |
kuchi_check_remake | Checks a shot list against the reference: reused brand text, assets taken from the reference, generated logos or UI, placeholders, timing gaps and drift. Returns passed and compare_frames. | Optional; useful to sanity-check a plan before handing it over. |
kuchi_list_components | Lists the copy-paste Remotion components (kinetic type, captions, UI screens, cursors, logos, transitions) with dependencies and shadcn install commands. | Rarely; useful to name components in the plan. |
kuchi_get_component | One component's full source, files, dependencies, install command and usage notes, split into parts when long. | Rarely. |
kuchi_send_feedback | Reports a failure or a missing capability to the maintainers. The only tool that records what the agent writes. | When a tool fails or misleads. |
search | The generic search tool some clients expect: returns up to 10 video ids, titles and URLs. | Used by apps that only call search and fetch. |
fetch | The generic fetch tool: one video by the id from search, with notes, links and metadata. | Used by apps that only call search and fetch. |
How the blueprint times your video
A reference video is measured shot by shot: where each shot starts and ends, how it enters and how much of the frame moves. kuchi_plan_remake maps those shots onto your target length, between 5 and 180 seconds, at the frame rate you will render (12 to 120 fps, the reference's own by default).
If your length is at least 75% of the reference's, every shot is kept and re-timed by the same factor (mode retimed). If it is much shorter, speeding everything up would wreck the rhythm, so the plan keeps an evenly spaced subset of shots, at least three, and re-times those instead (mode condensed). Either way every shot starts on a whole frame and lasts at least one, so start_frame and end_frame are exact integers a coding agent can place directly; no rounding happens later.
Each shot also carries compare_at_s: the moment in your video that matches the reference frame shown on the storyboard, so the side-by-side check compares like with like. precision says whether every shot was classified (analysed), some were (partly_analysed) or none were (timing_only); shots without a kind keep their timing and get placeholder content until ChatGPT classifies them. One shot from a blueprint:
{
"shot": 3,
"start_s": 4.2,
"end_s": 6.533,
"start_frame": 126,
"end_frame": 195,
"reference_shot": 3,
"reference_start_s": 3.9,
"reference_end_s": 6.07,
"compare_at_s": 5.1,
"enters": "transition",
"motion": "medium",
"kind": "ui_screen",
"reference_shows": "Product UI. Dashboard centered, cursor moves to a button",
"your_content": "Acme: \"Live sync\". A screenshot or recording of the user's product..."
}ChatGPT does not render the file. Ask it for the blueprint as JSON and give it to a coding agent such as Claude Code or Codex: each shot becomes a Remotion <Sequence> starting at start_frame and lasting end_frame − start_frame + 1 frames, inside a composition whose length is the blueprint's frames.
The 14 shot kinds ChatGPT classifies
While studying the storyboard, ChatGPT gives every reference shot one kind. The kind decides what of yours replaces it.
| Kind | In the reference | In your video |
|---|---|---|
title_card | Words on a plain or lightly textured background | A short line in your words |
kinetic_type | Words that are the motion: typing, sliding, scaling | Your headline, animated the same way |
logo | A brand mark on its own, usually a reveal | Your logo |
ui_screen | A whole product screen, flat to the camera | A screenshot or recording of your product |
ui_closeup | One part of a screen enlarged: a button, a field, a cursor | The matching detail of your UI |
device | A screen inside a phone, laptop or browser frame | Your screen in the same frame |
chart | A graph, counter or data graphic drawing in | Your real numbers, never invented ones |
character | A mascot, illustrated figure or avatar | Your own mascot, or a placeholder |
product_render | A physical product, photographed or in 3D | Your product photo or render |
live_action | Filmed footage of people or places | Your own footage, or a placeholder |
collage | Several images, cards or screens arranged together | Several of your screens |
abstract | Shapes, gradients, particles or light | The same motion in your colours |
end_card | The closing frame: logo with a call to action | Your logo and call to action |
blank | An empty frame held as a pause | The same pause, nothing on screen |
Entrances and motion
| Value | Meaning |
|---|---|
enters: cut | Hard cut: the whole picture changes in one frame. |
enters: transition | Animated transition: several frames of heavy change join two pictures (wipe, zoom, morph or dissolve). |
enters: hold | Same scene: motion resumes after the picture held still. |
enters: blank_black | Passes through an empty black frame. |
enters: blank_white | Passes through an empty white frame. |
motion: still | Under 0.4% of the frame moves. |
motion: low | Under 3%. |
motion: medium | Under 12%. |
motion: high | 12% or more. |
What the check catches before you ship
kuchi_check_remake takes the shot list the build used: each shot's start and end in seconds, the reference_shot it rebuilds, how it enters, every line of on-screen text, and where each asset came from (user, generated, stock, placeholder or reference). It returns passed, errors and warnings by shot, and compare_frames.
- Text that carries over the reference's brand, product names or slogans.
- Any asset taken from the reference video: footage, UI, characters or music.
- Generated logos or generated product UI, which should always come from you.
- Placeholders still on screen.
- Gaps or overlaps in the timeline, and shots whose length or entrance drifted from the reference shot they rebuild.
The check reads the shot list, not the file: the video never leaves the machine that renders it. The coding agent that builds the video runs the check again on its own shot list.
Troubleshooting ChatGPT
| Symptom | Fix |
|---|---|
| The plugin is created but never used | Turn it on in the chat before asking. |
| ChatGPT only searches and fetches | That surface only calls search and fetch. Use a chat where the full plugin is available, or ask it to call kuchi_suggest_templates by name. |
| Creation fails with an auth error | Choose No authentication; OAuth is not needed. |
| Your workspace hides the option | An admin must allow custom MCP servers. |
| The tool list is empty or the connection times out | Check that nothing on your network (a proxy, VPN or firewall) blocks HTTPS to kuchi.kalarislabs.com, and that the URL ends in /mcp. |
| A tool says the catalog is temporarily unavailable | The library is refreshing. Retry in a minute; nothing on your side needs to change. |
Launch video templates to start from
All launch videos →Components for the build
All components →ChatGPT and Kuchi: questions
How do I connect an MCP server to ChatGPT?
In ChatGPT open Plugins, select + → Add custom MCP server, name it Kuchi, paste https://kuchi.kalarislabs.com/mcp and choose No authentication.
Can ChatGPT render the launch video?
ChatGPT plans it: the template, the shot breakdown and a frame-timed blueprint with your content. A coding agent such as Codex renders the MP4 from that blueprint.
Is Kuchi free to use with ChatGPT?
Yes. The Kuchi MCP server is open: no account, no API key and no payment. Your own agent plan applies as usual.
Will my video copy the reference?
No. A remake borrows the reference's timing, motion and layout. Every logo, screen, line of copy and asset on screen comes from your product.







