Your imagination. Your creative copilot.

Dream it together.
Build it here.

Give your agent an idea. The Sogni Creative Agent Skill helps turn it into original art, then you bring your world to life in our guided builder. Start with one scene and two paths.

Make my agent brief Or build the full cinematic version, from an idea or your photos No coding agent? Let Sogni’s DeepSeek build it New to agent skills?

Planning here is free. You choose when to generate and how to pay.

YOU + YOUR AGENTFind the spark
A place, a little mystery, a reason to explore.
SOGNI CREATIVE AGENTMake the first scene
Original art, with your feedback at every step.
THE WORLD BUILDEROpen a path
Choose an object. Imagine what happens next.

The full cinematic version · open source

Build a world like The Dream Thread from an idea, or one like The Long White Cloud from your photos.

The Sogni Worlds Kit is the pipeline behind our showcase worlds, free on GitHub. Your own agent (Claude Code, Codex or Hermes) follows its instructions and does the work: it paints the places from your concept (or takes your photographs), the things you click carry you on through a door, a cloud or a waterfall, a choice can be fatal, figures hide in the pictures, and narration and music tell the story. You approve every film, then publish the world as your own website.

Places painted from an idea, or your photographs
Journeys caused by what you click
Fatal choices, endings and 3D collectibles
Narration, music, and your own site hosted anywhere

Get the kit on GitHub Play The Dream Thread Play The Long White Cloud

For a whole world we recommend Sogni Unlimited Pro: four MiniMax H3 films render at once, with twice the H3 fair-use capacity.

1
Get the kit
Needs Node 22.12+, ffmpeg and git.
2
Connect Sogni
Saves your API key, checks your plan and installs the Creative Agent Skill.
3
Hand it to your agent
Run node world new my-trip, open your agent in the folder and give it an idea:
Build a Sogni World from this concept: a lighthouse keeper’s last night on a rock in the Atlantic, in seven places. Two choices are fatal. Paint the places.
Or copy your photos into worlds/my-trip/photos/ and say:
Build a Sogni World from the photos in worlds/my-trip.
No coding agent? Let Sogni’s DeepSeek do this step.

No coding agent? · Sogni’s own LLMs do the agent’s work

Let Sogni’s DeepSeek build it for you.

In step 3, run node world agent my-trip instead of opening a coding agent. DeepSeek V4 Flash interviews you, plans every place, writes the direction for each film and screens every take; Qwen 3.6 finds the things you click. Your pictures, words and renders go to Sogni and nowhere else. The agent never approves a film: it can only reject one that is clearly broken, and everything it keeps waits for you on your review page.

Step	The agent	You
Interview	Asks what only you know: what the world is about, who is in it, each place’s story, narration, music.	Answer, or hand it a brief file.
Pictures	Copies your photos in story order, or paints the places from your concept when you have no photos.	Approve the concept.
Plan	Studies every still, writes each place’s title, narration, loop and the film behind each click, then checks its own work against the craft rules.	Approve the plan.
Render	Quotes first, renders, screens every take, and rewrites what is clearly broken.	Approve the quote.
Judge	Leaves a note on every take it keeps, then opens your review page.	Approve or reject every take.
Finish	Narration, music, 3D figures, the credit, a build.	Play it, then export it.

Cost. The LLM work for a three-place world is a few dozen calls and well under a dollar pay as you go; Sogni Unlimited covers it. Films cost exactly what they cost with any agent.
Time. Slower than a frontier coding agent: allow a few minutes per place for the plan when DeepSeek is busy, then the films render at their usual pace. Stop any time with Ctrl-C; the same command picks up where it left off.
Limits. It writes from what it sees and what you tell it, so tell it each place’s story. It cannot hear: listen to every take yourself. It is more literal than a frontier agent, so read the plan before you approve it.

01 / A small beginning

What kind of world is calling?

A sentence is plenty. We’ll put the creative recipe and next steps into a brief for your agent.

Borrow a spark

World name optionalWhat would you love to explore?

A setting, a character, or a curious “what if…” all work. Leave it open and your agent will help you choose.

The look and feel

How would you like to generate?Help me choosePlan first. Decide with my agent before generating.Sogni UnlimitedUse my plan wherever it covers the model. Ask before spending anything else.Pay as you goUse my Spark or SOGNI, within a budget we agree on first.

Sogni Unlimited · One plan, more possibilities

Give your next idea somewhere to go.

Try another opening. Give a character a new path. With Sogni Unlimited, the images and videos you make here come out of your plan instead of your Spark or SOGNI balance.

View Unlimited plans What’s included ↗︎

Plans have fair-use limits and a queue, and a few partner models cost extra. Prices and any free trial are shown at checkout.

Read the full brief before copying

Your ready-to-paste agent brief

Help me build an explorable world on Sogni World: https://worlds.sogni.ai/

MY WORLD
Working title: Help me find a memorable title
Idea: Help me choose an imaginative place, a reason to explore, and a satisfying little discovery.
Visual style: Warm, richly illustrated storybook art
First milestone: one starting scene and two connected destinations that I can explore in the web app. Keep the first version small and delightful.

SET UP TOGETHER
Use the official Sogni Creative Agent Skill: https://github.com/Sogni-AI/sogni-creative-agent-skill
If it is not installed, follow its current installation instructions for this agent. The personal-skill setup command is npx setup-sogni-agent-skill --version=latest (keep --version=latest; without it the installer puts in an old skill that lacks this world recipe). It asks questions, so either have me run it in my own terminal or run it with --yes --no-credentials and help me save my key. Avoid a duplicate personal registration if its plugin is already installed. After setup, verify with sogni-agent doctor --json and confirm success; if it reports that a newer version is available, update before planning. A skill installed during this conversation may not load until the agent restarts: read the installed SKILL.md directly, or ask me to restart and paste this brief again. Help me configure my own API key locally using the official instructions; never ask me to paste it into this brief or the World website. Use the same Sogni account in the agent and the website.
Read https://worlds.sogni.ai/how-this-was-made and the skill's references/interactive-worlds.md for the creative recipe, then verify current supported models and commands. Historical example model names are not an active model registry.

PAYMENT
Help me choose between Sogni Unlimited and Pay as you go. Check the account and explain the proposed payment before generating. Do not start a paid job until I choose.

PLAN BEFORE GENERATING
Propose a simple three-scene map: where we start, two interesting things to click, where each choice takes us, and a small discovery at the end. Write clear visitor-facing scene names and action labels. Make the layout easy to explore on a phone. Show me the plan and proposed first generation for approval before spending.
Start with one original landscape still. The builder makes paths only from an image whose edges are each 256 to 2560 pixels, at a common shape such as 16:9, 3:2 or 4:3, so do not upscale it to 4K. Leave the skill's safety filter on (no --no-filter): My Sogni images lists only filtered image jobs from the last 48 hours, so build soon after generating. Make clickable objects visible and distinct. Compose around the place; vary a recurring character's position, depth, and gaze. If I want the same character across scenes, use saved reference images, not their name alone. Preserve original image bytes and matching aspect ratios; never use screenshots or extracted video frames as source scenes.
Use currently supported Sogni models. The example recipe uses Krea 2 for places, an identity-preserving image edit for continuity, SAM 3 for object selection, and MiniMax FastH3 for travel between the original first and last stills. Let the web builder handle its supported path-generation workflow. Approve the first result together before generating more. Keep job receipts, prompts, references, and original files so interrupted work can resume without duplicate jobs.

BUILD ON SOGNI WORLD
Open https://worlds.sogni.ai/?view=create&mode=guided
Help me sign in myself and create a private world. Choose the original starting image from my recent Sogni images (eligible images from the same account in the last 48 hours), or upload the original file for review if it is unavailable. Uploaded images may need approval before they can start a world.
If you have browser tools, help operate the guided builder in my signed-in session, showing me paid actions before submitting them. Otherwise, give me the exact next click and file to choose. Do not read, extract, copy, or ask for browser cookies or JWTs. Sogni API keys support generation, but World authoring currently requires the signed-in browser; do not invent a direct World publishing API or claim an API key can write worlds.
In the builder, choose a visible object, describe the journey, generate and review the next scene and crossing, and repeat for the second path. Once both paths work, the builder can also turn an object into a moment (the scene comes alive where it stands) or a 3D collectible; offer these, priced, as optional extras. Describe physical travel rather than a dissolve between pictures. Keep the world private while we test every path and the way back on desktop and mobile, checking original image detail, motion, and sound.
Finish by giving me my private world link and a short checklist of what works and what still needs attention. Guide me through the available sharing or community-review controls only when I choose to share. Never make my world public automatically.

POLISH WHEN THE FIRST PATH WORKS
Use https://worlds.sogni.ai/how-this-was-made#polish-checklist as a finishing checklist. These are recommendations for manual or custom world builds, not a promise that the web builder can import finished videos or replace its canonical scene originals. Keep the first milestone small; discuss a custom build with the Sogni Worlds Kit (below) before expanding scope.
- Give the scene sound effects. Add music separately. Prompt for specific Foley: footsteps on gravel, a door latch, fabric rustling, water splashing. Ask for no music in the video and listen to the result; a prompt alone does not guarantee it. A separate score can flow across scenes without competing soundtracks.
- Master the sound for the finished mix. Listen with your intended music playing. Bring quiet Foley forward so the action still has impact, keep dialogue clear, and lower the score when someone speaks. Match levels between clips and leave headroom so louder effects do not clip or distort; avoid a blanket volume boost.
- Let the final video supply the display still. For a custom player, show the video’s own first or last frame between playback and the static scene. Browsers draw video and images through different color paths (on a Mac, video mid-tones come out brighter), so even a careful screenshot can pop. A one-frame clip cut from the same video file, with its color tags intact, is drawn exactly like the film. Check every incoming and outgoing path, in each picture setting. Keep the original generated images as your sources for future generations and object masks; endpoint frames are display-only.
- Tidy geometric selections. Keep organic detail. SAM 3 can return wonderfully intricate masks, but a wobbly door outline looks odd. Ask your agent to straighten and gently smooth doors, panels, and other geometric edges while keeping the selection on the object. Preserve the detailed contours of people, plants, hair, and fur. Show outlines on hover or focus during moving scenes, and check the touch targets on a phone.
- Make the click match the consequence. Watch the finished film before choosing its label and clickable object. Clicking a door should use that door; clicking a character should get a response. Give each journey a clear action and a payoff, then test the way back too.
- Keep characters recognizable and voices consistent. Reuse saved character reference images across scenes and an approved voice reference for recurring dialogue. A name or written description alone can cast someone new. Vary the character’s position and gaze so the world feels lived in.
- Watch the whole loop, including the seam. Living scenes need believable movement and audible atmosphere. Watch for objects appearing or disappearing, frozen characters, sudden light changes, and sound clicks where the loop repeats. Matching endpoint pictures alone does not prove a seamless loop.
- Play it like a first-time visitor. Review full motion and sound, then inspect faces, textures, edges, and gradients at native size. Try every path on desktop and a real phone, in both picture settings, with music and sound switched separately. Test slow loading and reduced motion, and keep your original files and approved takes safe while you polish.

SCOPE
The first milestone uses the live web builder's supported scenes, paths, moments and collectibles. The showcase's music score, living looped scenes, recurring cloned voices and rendered journeys back come from a custom build; do not promise that uploading a folder or a finished video imports those features into World. Show me any unsupported step before generating media the app cannot accept.
For that full cinematic version, use the open-source Sogni Worlds Kit: https://github.com/Sogni-AI/sogni-worlds-kit. Its AGENTS.md is the complete recipe and its `node world` commands plan, render, screen, narrate, score and build the world; it publishes as a static site I host myself, not as a Sogni World page. Start with: git clone https://github.com/Sogni-AI/sogni-worlds-kit && cd sogni-worlds-kit && npm install, then node world setup.

02 / Meet your creative copilot

An agent skill is a creative toolkit.

It teaches a local AI agent how to make images, video, and music on Sogni. Your brief tells it what to make and when to pause for you.

Open your agent

Use an agent that can run tools on your computer, such as Codex, Claude Code, or Hermes. Paste your brief and ask it to help you set up.

Already have the Sogni skill or plugin? Your agent will check it’s up to date, then you’re ready to plan.

Connect your Sogni account

Get an API key from the account menu on your Sogni dashboard ↗︎. Follow the skill’s setup to save it locally.

Use the same account here so your generated images are easy to find.

Make something you love

Your agent starts with a small map and one scene. Give feedback, approve the art, then bring that original image into the builder.

First scene missing the magic? Refine it before making the rest.

I’d like to install the skill myself

Open your local agent once, then run this in your terminal. The installer detects supported local agents and guides you through connecting Sogni. If you already installed the Sogni plugin, use its setup instructions to avoid installing the same skill twice.

After installation, check the connection. Look for "success": true before continuing.

A plain chat window that cannot run tools will help plan, but cannot run the local skill. Official setup for your agent ↗︎

03 / Make it explorable

Your first scene is a doorway.

Sign in, choose your original scene from your recent Sogni images, and create a private world. Pick an object, describe where it leads, then preview your new path. Your agent can guide you—or help operate the builder with browser tools while you review paid actions.

Need to upload a file instead? Use the original image; uploads may need review before you can build. New worlds stay private until you choose to share.

Bring my scene to the builder Explore the full creative recipe

Lessons from building these worlds

The finishing touches.

Start small, then polish. Use this checklist with your agent when you’re assembling a world by hand. Custom playback and audio work may need changes to your own player.

Open the polish checklist 8 practical tips

Give the scene sound effects. Add music separately.Prompt for specific Foley: footsteps on gravel, a door latch, fabric rustling, water splashing. Ask for no music in the video and listen to the result; a prompt alone does not guarantee it. A separate score can flow across scenes without competing soundtracks.
Master the sound for the finished mix.Listen with your intended music playing. Bring quiet Foley forward so the action still has impact, keep dialogue clear, and lower the score when someone speaks. Match levels between clips and leave headroom so louder effects do not clip or distort; avoid a blanket volume boost.
Let the final video supply the display still.For a custom player, show the video’s own first or last frame between playback and the static scene. Browsers draw video and images through different color paths (on a Mac, video mid-tones come out brighter), so even a careful screenshot can pop. A one-frame clip cut from the same video file, with its color tags intact, is drawn exactly like the film. Check every incoming and outgoing path, in each picture setting. Keep the original generated images as your sources for future generations and object masks; endpoint frames are display-only.
Tidy geometric selections. Keep organic detail.SAM 3 can return wonderfully intricate masks, but a wobbly door outline looks odd. Ask your agent to straighten and gently smooth doors, panels, and other geometric edges while keeping the selection on the object. Preserve the detailed contours of people, plants, hair, and fur. Show outlines on hover or focus during moving scenes, and check the touch targets on a phone.
Make the click match the consequence.Watch the finished film before choosing its label and clickable object. Clicking a door should use that door; clicking a character should get a response. Give each journey a clear action and a payoff, then test the way back too.
Keep characters recognizable and voices consistent.Reuse saved character reference images across scenes and an approved voice reference for recurring dialogue. A name or written description alone can cast someone new. Vary the character’s position and gaze so the world feels lived in.
Watch the whole loop, including the seam.Living scenes need believable movement and audible atmosphere. Watch for objects appearing or disappearing, frozen characters, sudden light changes, and sound clicks where the loop repeats. Matching endpoint pictures alone does not prove a seamless loop.
Play it like a first-time visitor.Review full motion and sound, then inspect faces, textures, edges, and gradients at native size. Try every path on desktop and a real phone, in both picture settings, with music and sound switched separately. Test slow loading and reduced motion, and keep your original files and approved takes safe while you polish.

Dream it together.Build it here.