sogni-ai/sogni-creative-agent-skill

Sogni Creative Agent Skill

Image, video, and music generation for AI agents — powered by Sogni AI’s decentralized GPU network.

---

Sogni Creative Agent Skill plugs into the agent runtime you already use — Claude Code, goose, OpenClaw, Hermes Agent, Manus AI, and others — and gives it production-quality image, video, and music generation through a single CLI: sogni-agent.

It ships four ways:

• a standalone Node.js CLI (sogni-agent)
• a skill source that any SKILL.md-aware agent can load
• a local MCP stdio server (sogni-agent-mcp) for goose, Claude Desktop, Codex, and other MCP clients
• a published OpenClaw plugin

With this skill, an agent can:

• generate with MOSS-Image SFT from captions, tags, or JSON layouts (-m moss_image_sft_int8_convrot), or the 8-step preset (-m moss_image_sft_dmd8_int8_convrot); both accept compatible Personal LoRAs. Optional --moss-rope-scale-x and --moss-rope-scale-y adjust spatial composition (0–100; neutral 1) while preserving output dimensions. See the model recipes.

• generate images from prompts and edit/restyle existing images

• select an object in an original image and return its SAM 3 mask (--segment; see the object selection guide)

• create videos from text, images, first/last frames, audio, or source video (LTX-2.5), with LTX-2.3 retained for voice ID, transition, and 10Eros workflows

• turn an image folder into a visually deduplicated, music-backed seamless loop with one plugin skill invocation

• generate instrumental music or full songs with lyrics

• run hosted creative workflows including storyboard-driven video

• save personas, preferences, and last-render state across sessions

• check balances, list models, and refine previous results

Fastest install: paste this repo’s GitHub URL into your agent and ask it to “install this skill”.

---

Table of Contents

• Quick Start
• Requirements
• Installation
  • Node CLI (default)
  • Claude Code plugin
  • OpenAI Codex CLI
  • Hermes Agent
  • goose
  • OpenClaw plugin
  • ChatGPT (Custom GPT)
  • Manus / other SKILL.md frameworks
  • Manual install from source
  • Host launchers
  • Verify your install
  • Upgrading safely from inside an agent
• Claude Desktop
  • Also works in OpenAI Codex
• Setup (Sogni API key)
• Usage
• CLI Reference
  • Common options
  • Quality presets
  • Recommended models
• Video Sizing & Aspect Ratios
• LTX-2.x Prompting Guide
• Photobooth (Face Transfer)
• Personas, Memory, and Personality
• Hosted API Modes
• Dynamic Prompt Variations
• Token Auto-Fallback
• Sogni Unlimited Subscription
• Error Reporting & Output
• For AI Agents
• Development
• License

---

Quick Start

1. Get a Sogni API key from dashboard.sogni.ai (open the account menu) and save it — see Setup.

2. Install (one command):

   npx setup-sogni-agent-skill --version=latest
   

   This auto-detects Claude Code, OpenAI Codex CLI, and Hermes Agent; installs the CLI globally; registers the skill into each detected runtime; prompts for your API key; and tells you how to request ChatGPT Custom-GPT setup instructions. (It does not configure OpenClaw — see the OpenClaw plugin section.)

   Prefer to do it manually? Install the CLI directly:

   npm install -g @sogni-ai/sogni-creative-agent-skill@latest
   sogni-agent --version
   

   Then point your agent runtime at this repository’s SKILL.md.

3. Verify the install:

   sogni-agent doctor
   

Then ask your agent to do something:

• “Generate an image of a sunset over mountains”
• “Edit this image to add a rainbow”
• “Use Sogni Loop Maker on my image folder”
• “Make a video of a cat playing piano”
• “Generate a 30 second synthwave product-launch theme”
• “Turn my selfie into James Bond using photobooth”
• “Refine the last image at higher quality”

---

Requirements

• Node.js ≥ 22.11.0
• Sogni API key (dashboard.sogni.ai)
• ffmpeg + ffprobe (optional) — required for local utilities such as --angles-360-video, --concat-videos, timestamped frame extraction, and --verify-video; --upscale-video needs ffprobe. Set FFMPEG_PATH / FFPROBE_PATH to override discovery (a leading ~ is expanded).
• macOS, Linux, or Windows

---

Installation

Node CLI (default)

For most agents and human users:

npm install -g @sogni-ai/sogni-creative-agent-skill@latest
sogni-agent --version

Then point your agent/runtime at this repository’s SKILL.md. When an install request is ambiguous, install the CLI and skill source together — that’s the supported default.

Claude Code plugin

The Claude Code plugin shells out to the sogni-agent CLI installed above, so both steps are required. From inside Claude Code, register the marketplace and install the plugin:

/plugin marketplace add Sogni-AI/sogni-creative-agent-skill
/plugin install sogni-creative-agent@sogni

The first command registers a sogni marketplace with one plugin entry (sogni-creative-agent) backed by a lean Claude-Code-focused plugin-skills/sogni-creative-agent/SKILL.md; the second installs the plugin into Claude Code. The full skill spec still lives at the repository root SKILL.md.

Pick one registration per machine. Install either this plugin or the personal skill that npx setup-sogni-agent-skill --version=latest writes to ~/.claude/skills/ — not both. With both installed, Claude Code lists two near-identical skills, which wastes context and makes skill selection ambiguous.

OpenAI Codex CLI

For the namespaced Sogni skills, register the marketplace and install the plugin:

codex plugin marketplace add Sogni-AI/sogni-creative-agent-skill
codex plugin add sogni-creative-agent@sogni

Start a new Codex session, then invoke Loop Maker directly:

$sogni-creative-agent:loop-maker ./images

In the Codex app, type @, choose Sogni Loop Maker, and select its starter prompt.

For local development, replace the GitHub shorthand in codex plugin marketplace add with the absolute path to this repository.

Alternatively, the npx installer writes the full monolithic skill to ~/.codex/skills/sogni-creative-agent-skill/, which Codex discovers automatically:

npx setup-sogni-agent-skill --version=latest --only=codex

Start Codex once before running the installer so ~/.codex/ exists. If the selected local runtime is not detected, setup exits before installing anything.

Restart Codex (or start a new session) and ask it to “generate an image of a sunset” — the skill shells out to the globally installed sogni-agent. To remove this personal-skill install later: npx setup-sogni-agent-skill --uninstall --only=codex.

Hermes Agent

Hermes Agent can install the dedicated, security-scannable bundle directly from its Skills Hub (via skills.sh):

hermes skills install skills-sh/sogni-ai/sogni-creative-agent-skill/sogni-creative-agent-skill
npm install -g @sogni-ai/sogni-creative-agent-skill@latest
sogni-agent-hermes doctor

Then /reset the Hermes session so it picks up the new skill. The Hub bundle lives at .agents/skills/sogni-creative-agent-skill/ and contains only the agent instructions and the references they load, keeping the CLI runtime and development files outside Hermes’ skill security scan.

For a combined CLI + personal-skill setup, the npx installer still places the skill at ~/.hermes/skills/media/sogni-creative-agent-skill/:

npx setup-sogni-agent-skill --version=latest --only=hermes

Start Hermes once before running the installer so ~/.hermes/ exists. If the selected local runtime is not detected, setup exits before installing anything.

goose

goose supports both Agent Skills and MCP extensions. Pick one integration per installation so goose does not receive duplicate Sogni instructions and tools.

To install the skill globally:

npx skills add Sogni-AI/sogni-creative-agent-skill --global --agent goose --skill sogni-creative-agent-skill --yes
npm install -g @sogni-ai/sogni-creative-agent-skill@latest
sogni-agent-goose doctor

The skill is discovered from goose’s global skill directories and uses the sogni-agent-goose launcher for host attribution.

To run the MCP extension directly instead:

goose session --with-extension "npx -y --package @sogni-ai/sogni-creative-agent-skill@latest sogni-agent-mcp"

The same command can be added permanently from goose configure as a Command-line Extension. Once the public listing is accepted, it can also be installed from the goose Extensions directory. The MCP package is self-contained; it reads SOGNI_API_KEY from the extension environment or the existing ~/.config/sogni/credentials file.

OpenClaw plugin

The skill is published on ClawHub by the sogni-ai publisher. Specify the publisher because more than one listing uses this skill name:

npx -y clawhub@latest install @sogni-ai/sogni-creative-agent-skill

Through 3.54.2 the listing was @fishmongr/sogni-creative-agent-skill. ClawHub no longer finds that name, so an install made from it does not update; install again with the command above.

To install as a code plugin instead, use OpenClaw’s npm: source prefix (the npm package is scoped, so a bare openclaw plugins install sogni-creative-agent-skill will not resolve it):

openclaw plugins install npm:@sogni-ai/sogni-creative-agent-skill

The installed plugin loads its behavior from SKILL.md via openclaw.plugin.json. The npx setup-sogni-agent-skill --version=latest installer does not configure OpenClaw — use the command above (or the local-link flow below) instead.

API key under OpenClaw: the plugin config holds non-secret defaults only (models, timeouts, paths) — it does not carry your API key. Provide SOGNI_API_KEY via the environment the OpenClaw gateway passes to the CLI, or save it to ~/.config/sogni/credentials (SOGNI_API_KEY=). This keeps your key out of plugin config files.

For a local checkout that you want to update continuously, link the minimal OpenClaw surface (.openclaw-link/) — not the repository root, which contains development tests that OpenClaw correctly blocks during plugin safety scanning:

cd /path/to/sogni-creative-agent-skill
npm install
npm link
npm run openclaw:sync
openclaw plugins install -l "$PWD/.openclaw-link"
openclaw gateway restart

To update the linked install later:

cd /path/to/sogni-creative-agent-skill
git pull --ff-only
npm install
npm link
npm run openclaw:sync
openclaw gateway restart

The generated .openclaw-link/ directory is only for OpenClaw; Hermes, Manus, and other skill-based agents should continue using the root SKILL.md.

OpenClaw configuration

When loaded through OpenClaw, this skill reads plugin defaults from OpenClaw config; CLI flags always override them. The supported config schema is defined in openclaw.plugin.json and includes default models, video workflow models, hosted API defaults (apiBaseUrl, defaultLlmModel, defaultTaskProfile, defaultApiMaxTokens, defaultApiThinking, defaultApiToolMode, workflow cost defaults), token type, seed strategy, timeouts, and media paths. If your OpenClaw config lives elsewhere, set OPENCLAW_CONFIG_PATH.

ChatGPT (Custom GPT)

Run npx setup-sogni-agent-skill --only=chatgpt --version=latest to print step-by-step instructions for creating a ChatGPT Custom GPT whose Instructions embed this skill. Note that ChatGPT cannot run the local CLI; the Custom GPT path covers prompt-side behavior only.

Manus / other SKILL.md frameworks

Point the agent at this repository’s SKILL.md for behavior guidance and llm.txt for install/setup help. The agent should invoke the globally installed sogni-agent CLI by default.

Manual install from source

gh repo clone Sogni-AI/sogni-creative-agent-skill
cd sogni-creative-agent-skill
npm install

Host launchers

The package installs one launcher shim per host alongside sogni-agent. Each shim is the same CLI with the same flags and output — it only tags the request with the agent framework that ran it, so Sogni can tell a Codex render from a Claude Code render:

The Claude Code and Codex plugin surfaces already pin their launcher, and OpenClaw is detected automatically. Goose and Hermes skill installs should use their matching launchers wherever the docs say sogni-agent. Everything works normally through bare sogni-agent — the render is just attributed to unknown. Falling back to sogni-agent is always safe if a shim is not on PATH.

Verify your install

Every install path above ends the same way — run the built-in health check:

sogni-agent doctor

It verifies the Node version, API credentials (and their file permissions), config-dir writability, ffmpeg availability, live authentication, and whether a newer version is available. sogni-agent doctor --json emits the same checks for agents. If anything is marked ✗, the detail line says exactly how to fix it.

Upgrading safely from inside an agent

When upgrading from inside an agent runtime, prefer direct package-manager or existing-checkout commands. Avoid asking the agent to build a clone-or-pull shell bootstrap script with set -e, bash -c, sh -c, or an inline repository URL — some sandboxes correctly route those through approval and the install will stall.

For a global CLI:

npm install -g @sogni-ai/sogni-creative-agent-skill@latest
sogni-agent --version

For an existing local checkout:

DEST="$HOME/Documents/git/sogni/sogni-creative-agent-skill"
git -C "$DEST" pull --ff-only
npm --prefix "$DEST" install

If the checkout is missing, use the npm install path above or explicitly approve a clone.

---

Claude Desktop

Claude Desktop can’t run skills against your local files, so Sogni ships as a local MCP server instead. Two ways to install:

Recommended — one command (also installs the CLI, saves your API key, and offers to install ffmpeg):

npx setup-sogni-agent-skill --version=latest

This registers the Sogni tools in claude_desktop_config.json. Fully quit and reopen Claude Desktop afterwards. Generated images display inline in the chat automatically.

Manual — drag-and-drop bundle: download sogni-creative-agent.mcpb from the GitHub Releases page and drop it onto Claude Desktop’s Settings → Extensions page. You’ll be prompted for your Sogni API key (stored in the OS keychain) unless you’ve already run the installer.

Don’t use both — you’d get duplicate Sogni tools. The extension wraps the same globally installed sogni-agent CLI used by Claude Code, so personas, memories, and credentials are shared.

Video/audio editing features need ffmpeg on your machine; the npx installer offers to install it for you.

Also works in OpenAI Codex

The same local MCP server runs in OpenAI Codex — the Codex CLI and IDE extension read MCP servers from ~/.codex/config.toml. With the CLI installed globally (npm i -g @sogni-ai/sogni-creative-agent-skill), register it and start a new Codex session:

codex mcp add sogni-creative-agent -- node "$(npm root -g)/@sogni-ai/sogni-creative-agent-skill/desktop-extension/server/index.mjs"

Or write the ~/.codex/config.toml entry yourself (use the absolute path from npm root -g):

[mcp_servers.sogni-creative-agent]
command = "node"
args = ["/opt/homebrew/lib/node_modules/@sogni-ai/sogni-creative-agent-skill/desktop-extension/server/index.mjs"]

Codex also runs the Skill natively (see OpenAI Codex CLI) — pick one integration per machine; running both gives Codex duplicate Sogni tools. The server finds your globally installed sogni-agent, ffmpeg, and saved credentials on its own, so no env block is needed.

---

Setup (Sogni API key)

1. Get your API key from dashboard.sogni.ai (open the account menu).

2. Save it to a credentials file:

   mkdir -p ~/.config/sogni
   cat > ~/.config/sogni/credentials << 'EOF'
   SOGNI_API_KEY=your_api_key
   EOF
   chmod 600 ~/.config/sogni/credentials
   

You can also skip the file and export SOGNI_API_KEY in your environment.

Filesystem path overrides

Defaults live under ~/.config/sogni/ for credentials, last-render metadata, personas, memories, and personality.

Running several agents at once (Claude Code, Codex, OpenCode, hermes, …) works out of the box: each process leases its own stable app ID from a persistent slot pool in ~/.config/sogni/app-ids/, so concurrent runs never fight over one socket identity (SWITCH_CONNECTION 4015) and routine runs do not create unnecessary IDs (error 4061). Leases are released on exit and reclaimed automatically if a process dies. Long-lived daemons should pin their own SOGNI_APP_ID to stay out of the shared pool, and each harness can set its own SOGNI_LAST_RENDER_PATH so --last stays per-tool. Override individual paths with:

---

Usage

Claude Code plugin users can launch the complete folder-to-loop workflow with one skill invocation:

/sogni-creative-agent:loop-maker ./images
/sogni-creative-agent:loop-maker ./images start=cover.jpg music="subtle tropical cyberpunk electronica" output=launch-loop.mp4

Codex CLI and IDE users can invoke the same bundled skill with $:

$sogni-creative-agent:loop-maker ./images
$sogni-creative-agent:loop-maker ./images start=cover.jpg music="subtle tropical cyberpunk electronica" output=launch-loop.mp4

The skill visually removes repeated concepts from the active sequence, renders one direct LTX first-frame/last-frame clip per image pair, closes the loop, generates a soundtrack longer than the picture, and verifies anchors, interior motion, streams, and full-file decoding before delivery. Original source images are preserved. The default stack is Sogni plus the bundled FFmpeg wrappers; HyperFrames and Remotion are optional only for explicitly requested text, overlays, or compositor effects. True 360 novel-view synthesis is deliberately excluded from this workflow because a direct single-image LTX prompt does not reliably create camera orbit geometry.

Image output supports --output-format png, jpg, or webp. For worker images, --no-prompt-metadata omits embedded prompt and generation settings. JSON output includes per-image results with nsfwDetected and nsfwSources; a label alone does not mean the optional content filter withheld the image.

# Image generation
sogni-agent -Q hq -o dragon.png "a dragon eating tacos"

# Edit an image
sogni-agent -c subject.jpg "add a neon cyberpunk glow"

# Photobooth face transfer
sogni-agent --photobooth --ref face.jpg "80s fashion portrait"

# Text-to-video (t2v) with native dialogue
sogni-agent --video 'A narrator says "welcome to the story" as ocean waves crash'

# Short-side resolution targeting (preserves the inherited aspect ratio)
sogni-agent --video --target-resolution 768 \
  "A calm cinematic shot of lanterns drifting across a night lake"

# Seedance 2.0 4K (4-15s vendor video path with native audio)
sogni-agent --video -m seedance2 --target-resolution 2160 --duration 8 \
  "A polished product reveal with native ambient sound"

# Seedance multimodal context with public HTTPS references
sogni-agent --video -m seedance2 --workflow t2v \
  --ref https://cdn.example.com/product.png \
  --ref-video https://cdn.example.com/motion.mp4 \
  --ref-audio https://cdn.example.com/music.m4a \
  "Use @Image1 for product identity, @Video1 for camera movement, and @Audio1 for music rhythm"

# Alibaba Wan 3 unified video (2-30s, fixed 30fps, native audio, 480P/720P/1080P; 720P by default)
sogni-agent --video -m wan3 --target-resolution 1080 --duration 8 \
  'A presenter walks through a detailed studio and says "Welcome."'
sogni-agent --video -m wan3 --ref first.png --ref-end last.png \
  "Move smoothly between the supplied endpoint frames"
sogni-agent --video -m wan3 --workflow ia2v --ref presenter.png --ref-audio dialogue.mp3 \
  "Use Image 1 for the presenter and Audio 1 for the performance"
sogni-agent --video -m wan3 --workflow r2v --ref-video motion.mp4 \
  "Use Video 1 as loose motion and pacing guidance for a new rainy-night scene"
sogni-agent --video -m wan3 --smart-duration --wan3-ratio 9:16 --no-expand-prompt \
  --reference-file-url https://cdn.example.com/brief.pdf \
  "Use the supplied brief exactly"

# Seedance 2.5 Uncensored: every Seedance 2.5 mode and limit under its own model.
# Accept its one-time likeness and consent agreement in the Sogni app first
# (until then jobs fail with error 4103, which is not retryable).
sogni-agent --video -m seedance2-5-uncensored --target-resolution 1080 --duration 8 \
  "A slow dolly through a rain-soaked neon street at night"

# Seedance 2.0 Mini Uncensored: every Seedance 2.0 Mini mode and limit (480p/720p,
# 4-15s) under its own model. It shares the Seedance 2.5 Uncensored agreement.
sogni-agent --video -m seedance2-mini-uncensored --target-resolution 720 --duration 8 \
  "A slow dolly through a rain-soaked neon street at night"

# Wan 3.0 Enhanced through MuleRouter (smart/fixed 2-30s, 30fps, native audio)
sogni-agent --video -m wan3-enhanced --target-resolution 1080 --duration 8 \
  --wan3-ratio 16:9 'A presenter says "Welcome." in a detailed studio'
sogni-agent --video -m wan3-enhanced --workflow i2v --ref first.png --ref-end last.png \
  -c wardrobe.png "Keep the endpoint frames and use Image 1 for wardrobe detail"

# MiniMax H3 Standard, 8-step Balanced, 4-step LightX2V Turbo, FastH3 Turbo, and FastH3 and Ref2VA Two-Stage (1080p/2K) video
sogni-agent --video -m minimax-h3 --duration 10 ""
sogni-agent --video -m minimax-h3-i2v --ref first.png --duration 8 ""
sogni-agent --video -m minimax-h3-r2v --ref identity.png -c wardrobe.png \
  --ref-video motion.mp4 --ref-audio voice.m4a \
  ""
sogni-agent --video -m minimax-h3-balanced --duration 8 ""
sogni-agent --video -m minimax-h3-flf2v-balanced --ref first.png --ref-end last.png --duration 8 ""
sogni-agent --video -m minimax-h3-turbo --duration 8 ""
sogni-agent --video -m minimax-h3-turbo --sampler sa_solver --duration 8 ""
sogni-agent --video -m minimax-h3-flf2v-turbo --ref first.png --ref-end last.png --duration 8 ""
sogni-agent --video -m minimax-h3-fasth3-turbo --duration 8 ""
sogni-agent --video -m minimax-h3-fasth3-i2v-turbo --ref first.png --duration 8 ""
sogni-agent --video -m minimax-h3-fasth3-flf2v-turbo --ref first.png --ref-end last.png --duration 8 ""
sogni-agent --video -m minimax-h3-fasth3-turbo-2stage --duration 8 ""   # 2K: FastH3 Two-Stage, delivered at 2x the canvas
sogni-agent --video -m minimax-h3-fasth3-turbo-2stage --target-resolution 1080 --duration 8 ""   # 1080p: 960x544 delivered at 1920x1088
sogni-agent --video -m minimax-h3-r2v-2stage --ref identity.png -c wardrobe.png ""   # 2K Ref2VA Two-Stage (20 steps)
sogni-agent --video -m minimax-h3-r2v-balanced-2stage --target-resolution 1080 --ref identity.png ""   # 1080p Balanced Ref2VA Two-Stage (8 steps)

# MiniMax H3 FastH3 audio-to-video: your voice or song drives the clip and is its soundtrack
sogni-agent --video -m minimax-h3-fasth3-ia2v-turbo --ref portrait.png --ref-audio voice.m4a --duration 8 ""
sogni-agent --video -m minimax-h3-fasth3-flfa2v-turbo --ref first.png --ref-end last.png --ref-audio song.mp3 --duration 12 ""
sogni-agent --video -m minimax-h3-fasth3-a2v-turbo --ref-audio song.mp3 --audio-start 30 --duration 15 ""

# MiniMax H3 keyframes: the clip passes through extra images at chosen times (repeat --keyframe, up to 8)
sogni-agent --video -m minimax-h3-fasth3-flf2v-turbo --ref first.png --ref-end last.png \
  --keyframe [email protected] --keyframe wave.png@6 --duration 8 \
  " and describes it at its time>"

# Image-to-video (i2v; defaults to wan_v2.2-14b-fp8_i2v_lightx2v)
sogni-agent --video --ref cat.jpg "gentle camera pan"

# Animate two images together (first frame → last frame; defaults to
# ltx25-22b-int8_i2v_distilled using the standard first/last-frame template)
sogni-agent --video --ref first.png --ref-end last.png \
  "the opening frame flows smoothly into the final frame"

# Image+audio-to-video (auto-routes to LTX-2.5 ia2v)
sogni-agent --video --ref cover.jpg --ref-audio song.mp3 \
  "music video with synchronized motion"

# Pixal3D textured GLB from one original image (no prompt)
sogni-agent --image-to-3d object.png --mesh-faces 30000 -o object.glb

# Pixal3D multi-view: the front plus any of the subject's own left side, back and
# right side (left view = its own left side faces the camera, so it faces screen-left)
sogni-agent --image-to-3d front.png --left-view left.png --back-view back.png --right-view right.png -o object.glb

# BiRefNet transparent PNG (add --matte for the soft mask)
sogni-agent --remove-background original.png -o cutout.png

# MiniMax Music 3 (the default music model) and Qwen3-TTS studio speech
sogni-agent --music --duration 30 -o score.mp3 "instrumental orchestral theme, 90 BPM, D minor"
sogni-agent --speech --speech-voice ryan -o speech.wav "Welcome to the story."

# Voice cloning from a 3–30s recording, or voice design from a description
sogni-agent --speech --speech-mode clone --voice-reference voice.wav -o clone.wav "A new line in this voice."
sogni-agent --speech --speech-mode design --voice-description "Warm, low-pitched storyteller" -o designed.wav "The journey begins."

# Direct music generation (MiniMax Music 3; tempo and key go in the prompt)
sogni-agent --music --duration 60 \
  "uplifting cinematic synthwave theme for a product launch, 110 BPM, A minor"

# Song with lyrics (plain section tags, one per line)
sogni-agent --music --lyrics $'[Verse]\nRise with the morning light\n[Chorus]\nWe run' \
  --output-format mp3 "bright indie pop chorus, 128 BPM, C major"

# ACE-Step 1.5 XL Turbo by name: a cheap draft with exact BPM/key, up to 600s
sogni-agent --music -m turbo --bpm 128 --keyscale "C major" "bright indie pop chorus"

# LTX-2.3 voice identity / persona
sogni-agent --video --reference-audio-identity voice.webm \
  'NARRATOR: "This is my voice."'

# Hosted chat with Sogni creative-agent tools (/v1/chat/completions)
sogni-agent --api-chat \
  "Create a 4-shot product video concept for a red sneaker"

# Hosted chat with image vision plus media-reference metadata
sogni-agent --api-chat --ref product.jpg \
  "Turn this into a launch poster and describe the edit plan"

# Hosted chat controls and model discovery
sogni-agent --api-chat --task-profile reasoning --no-thinking \
  "Plan a concise multi-step product launch workflow"
sogni-agent --list-api-models

# Live Supernet media model discovery
sogni-agent --list-models
sogni-agent --search-models spicy
sogni-agent --list-models --model-media image --model-tag uncensored

# Live LoRA catalog discovery (IDs, strength ranges, bipolar directions)
sogni-agent --list-loras --lora-catalog-model krea2_turbo_fp8_scaled
sogni-agent --search-loras lighting
sogni-agent --json --list-loras --lora-category character

# Durable hosted chat run with SSE progress events
SOGNI_SKILL_USE_SDK_TRANSPORT=1 sogni-agent --durable-chat \
  "Create a product launch storyboard and render the first hero image"

# Durable hosted workflow (/v1/creative-agent/workflows)
sogni-agent --api-workflow \
  --video-prompt "The camera slowly pushes in as the sketch comes alive" \
  "A graphite robot sketch on a drafting table"

# Durable workflow with a media reference and a cost ceiling
sogni-agent --api-workflow --ref https://cdn.example.com/sketch.png \
  --workflow-max-cost 25 --confirm-cost \
  --video-prompt "The camera slowly pushes in as the sketch comes alive" \
  "Animate the referenced sketch"

# Exact durable workflow input
sogni-agent --api-workflow --workflow-input @workflow.json

# Storyline -> GPT Image 2.5 Sunburst storyboard sheet -> Seedance 2.5 1080p video
sogni-agent --api-workflow storyboard-video --storyboard-frames 6 --duration 12 -Q hq \
  "Create a 9:16 bakery launch video with a neon street-window reveal"

# Sogni Intelligence replay records
sogni-agent --list-replays 20
sogni-agent --get-replay run_abc123 --json

# Opt in to SDK transport for hosted operations (durable workflows + chat).
# Validates restEndpoint/socketEndpoint via the skill's SSRF guard, then
# calls the SDK workflow/chat methods directly.
# Falls back to the legacy SSRF-validated fetch path when the env is unset.
export SOGNI_SKILL_USE_SDK_TRANSPORT=1
sogni-agent --api-workflow storyboard-video "10s neon city flyover"

# Local segment + concat with external soundtrack
sogni-agent --video --workflow v2v --ref-video dance.mp4 \
  --video-start 10 --duration 8 --controlnet-name pose -o ./clip-2.mp4 \
  "robot dancing"
sogni-agent --concat-videos ./final.mp4 ./clip-1.mp4 ./clip-2.mp4 \
  --concat-audio song.mp3 --concat-audio-start 0

# Balances and help
sogni-agent --balance
sogni-agent --help

Prefer .webm, .m4a, or .mp3 voice clips. Local .wav clips are normalized to .m4a before upload when ffmpeg is available.

For local multi-clip workflows, use the built-in FFmpeg wrappers (--video-start, --audio-start, --audio-duration, --concat-videos, --concat-audio) over raw shell commands — they produce safer, more reproducible results.

---

CLI Reference

Run sogni-agent --help for the full CLI. Below are the options and tables most agents and users reach for first.

Common options

Quality presets

Skip remembering model IDs — --quality / -Q selects the right model, steps, and dimensions for image generation:

Explicit --model overrides the preset’s model. Explicit -w/-h overrides dimensions.

Recommended models

Prefer -Q fast|hq|pro for images and automatic workflow routing for video. Pass -m only when you need a specific model family.

gpt-image-2 supports flexible OpenAI image sizes up to 3840 px on either edge, max 3:1 aspect ratio, and total pixels from 655,360 to 8,294,400; the API snaps dimensions to valid multiples of 16. For image editing with gpt-image-2, you can pass up to 16 context images. For likeness-preserving edits of a referenced person or character, agents default to Krea 2 Identity Edit (krea2_identity_edit_v1_2) unless you explicitly choose another model. It and Dark Beast Krea 2 Identity Edit (dark_beast_krea2_identity_edit_v1_2) use -c/--context, accept 1-2 references at 512-2048 px, and leave execution defaults to the current model tier.

For Pixal3D mesh budgets, BiRefNet cutouts/mattes, Music 3 section tags, and all Qwen3-TTS modes, read the model guide. Speech uses exact scripts without music duration or sampler controls; clone mode checks 3–30s recordings with ffprobe. Audio batches save every output with numbered filenames.

Music generation uses --music and outputs mp3 by default. --audio remains the video-reference alias for --ref-audio; use --music or --generate-music for direct audio-only generation.

Seedance 2.5 loose-reference operations use an explicit task contract. reference accepts any loose image/video/audio set (including audio-only); edit and extend require @Video1 through --ref-video and reject first/last-frame anchors. Edit inherits the source aspect ratio and requires its source duration, while extend inherits the ratio and uses the requested continuation duration:

sogni-agent --video -m seedance2-5 --seedance-task-type reference --ref-audio voice.m4a "Use @Audio1 to guide a new performance"
sogni-agent --video -m seedance2-5-v2v --seedance-task-type edit --ref-video source.mp4 --duration 8 --target-resolution 720 "Edit @Video1 while preserving its subject and timing"
sogni-agent --video -m seedance2-5-v2v --seedance-task-type extend --ref-video source.mp4 --duration 8 --target-resolution 720 "Extend @Video1 after its ending"

---

Video Sizing & Aspect Ratios

• WAN 2.2 models use dimensions divisible by 16, each side 480–1536 px, and at most 1,048,576 pixels per frame (1024×1024), so 1280×720, 720×1280, and 1024×1024 fit but 1536×864 and 1080p do not (as of 2026-09). The CLI refuses an explicit -w/-h or --target-resolution above that instead of sending it; sizes it chooses itself are fitted. (Since 2026-09-26 the network itself scales an oversized WAN 2.2 request down to about 1 MP and refuses only a side below 480, error 4101.)
• Wan 3 (wan3.0-video) is a Premium Spark Alibaba model with fixed 30 fps, fixed or smart 2–30 s output, native audio, 480P/720P/1080P tiers, and adaptive, 16:9, 4:3, 1:1, 3:4, or 9:16 output. It accepts one first frame plus an optional last frame, or loose references up to 10 images, 5 videos, and 5 audio clips. It can also use one public document (PDF/Office/text/Markdown/Keynote/Pages/Numbers, up to 100 MB; PDF/DOCX/DOC/PPTX/PPT/Keynote/Pages up to 50 pages) or one public webpage; file and link context are mutually exclusive and cannot be combined with first/last anchors. Each video/audio reference is 1–15 s and each modality’s clips total at most 15 s. Video references condition a new generation; Alibaba exposes no edit/extend task mode or task-type field, so use a dedicated video-to-video model when source-preserving editing is required. Provider prompt expansion defaults on; --no-expand-prompt preserves the exact prompt. Watermark defaults off. Send no negative prompt, steps, guidance, sampler, scheduler, ControlNet, or mask. Sogni retail rates before subscription discounts are $0.065/$0.13/$0.26 per output second at 480p/720p/1080p (as of 2026-08). The CLI renders 720P unless --target-resolution 480 or 1080 is given, sending Alibaba’s exact size for that tier and ratio; image-to-video keeps the first frame’s shape with its short side at the tier.
• Wan 3.0 Enhanced (wan3.0-spicy-video; MuleRouter provider ID w3.0-video; aliases wan3-enhanced, wan3-spicy) runs at fixed 30 fps with fixed or smart 2–30 s output, native audio, prompt expansion, 480P/720P/1080P, and adaptive, 16:9, 9:16, 1:1, 4:3, or 3:4. It accepts either first/last-frame anchors or up to 10 image, 5 video, and 5 audio loose references; the two modes cannot be mixed. It has no document/web context, watermark, negative prompt, source-video edit, or extend mode. Standard PAYG is $0.08/$0.16/$0.32 per output second at 480P/720P/1080P (as of 2026-09). Existing paid subscribers receive a one-time $10 Unlimited / $20 Unlimited Pro model credit through September 1, 2026 00:00 UTC (August 31 at 5:00 PM PT); this is a launch credit, not subscription coverage.
• MiniMax H3 Standard, Balanced, LightX2V Turbo, and FastH3 Turbo use a 32 px grid, fixed 24 fps, native 32 kHz stereo audio, 124–362 frames (124 + n×17, i.e. 5.17–15.08 s), and a 1,032,192-pixel render cap; Ref2VA Turbo’s cap is 522,240 pixels, its own 960×544 canvas, and the CLI fits larger requests inside it. Standard, Balanced, LightX2V FL2VA Turbo, and FastH3 normally default to 1344×768 or 768×1344; Ref2VA Turbo defaults to 960×544. Standard uses 20 steps; Balanced uses fixed 8-step Euler/simple acceleration, with LightX2V for FL2VA and Larry v4 for Ref2VA; both Turbo engines use 4 steps. FL2VA H3 Turbo defaults to er_sde on Socket and permits euler, er_sde, or sa_solver; Ref2VA Turbo and the separate FastVideo VSA FastH3 engine use Euler/simple only. The CLI omits the sampler unless --sampler is passed. FastH3 is about 2x faster than LightX2V Turbo and up to 6x faster than Standard for comparable 768p, 15-second requests; it supports T2V/I2V/FLF2V, never R2V. Guidance 1 is fixed and there is no negative-prompt input. Every Ref2VA tier accepts up to 9 images, 3 videos, and 3 audios, capped at 12 files; at least one visual reference is required. FastH3 keeps the FastVideo engine when an H3 LoRA is attached: base jobs require 23 GB and jobs with an H3 LoRA require 32 GB. The other FL2VA/Balanced/Turbo and image-only R2V paths require 32 GB-class workers, while video-conditioned R2V requires above 40 GB. See references/video-prompting.md § MiniMax H3 Prompting for the exact contracts.
• LTX family (ltx2-*, ltx23-*, ltx25-*) uses dimensions divisible by 64. The current wrapper caps non-WAN video dimensions at 2048 px on the long side.
• Seedance runs at fixed 24 fps. The 2.0 family (seedance2, seedance2-mini, Seedance 2.0 Mini Uncensored seedance2-mini-uncensored, seedance2-fast) supports 4–15 s durations; full seedance2 supports native 4K via --target-resolution 2160 while seedance2-mini, seedance2-mini-uncensored and seedance2-fast remain capped to the 720p lower-resolution path. seedance2-5 and Seedance 2.5 Uncensored (seedance2-5-uncensored) render 4–30 s single clips at 480p/720p/1080p (no 4K), with --output-format mov and --return-last-frame for the final frame as an image. Other default/WAN paths support up to 10 s; LTX and WAN animate workflows support up to 20 s.
• For spoken dialogue, budget roughly 3 words per second plus about 1 second for each meaningful acting beat or pause. Keep quoted speech under the model’s hard per-clip word budget.
• The script auto-normalizes video sizes to satisfy these constraints.
• Use --target-resolution for bare resolution requests like “720p” — it targets the short side and preserves the inherited aspect ratio.
• Natural-language aspect requests like “portrait”, “square”, “16:9”, or “9:16” are inferred when width/height aren’t explicitly set. Combined requests like “720p 9:16” keep the requested short side while applying the requested shape.
• For i2v (and any workflow using --ref / --ref-end), the client wrapper resizes the reference image with strict aspect-fit (fit: inside) and uses the resized dimensions as the final video size. Because that resize uses rounding, a “valid” requested size can still produce an invalid final size (example: 1024×1536 requested, but ref becomes 1024×1535). sogni-agent detects this for local refs and auto-adjusts to a nearby safe size.
• LTX-2.5 first/last-frame default: --ref + --ref-end with no -m uses ltx25-22b-int8_i2v_distilled and the dedicated FLF workflow template. It shares the I2V public model ID and does not attach the LTX-2.3 transition LoRA. Pin ltx23-22b-fp8_i2v_distilled explicitly only when the legacy transition-LoRA behavior is required.
• Private mature-theme creativity: optional uncensored LTX-2.3 video models are available for adults who explicitly want them. They remain opt-in and are not part of ordinary model recommendations; the agent loads their specialized guidance only for a relevant request.
• Pass --strict-size to fail instead — the script will print a suggested size.

V2V defaults mirror Sogni Chat workflow tuning: canny, pose, and depth use ControlNet strength 0.85 with detailer assist; detailer uses strength 1.0. LTX 2.5 pose requires both --ref-video for the motion/pose sequence and --ref for the subject appearance; explicit LTX 2.3 rollback keeps its existing optional-image behavior. Use -m seedance2-v2v for Seedance V2V without ControlNet. Seedance accepts public HTTPS image, video, and audio references that pass CLI URL safety checks; localhost and private-network URLs are rejected before forwarding. The 2.0 family requires audio references to be paired with an image or video; Seedance 2.5 reference mode permits audio-only input.

The LTX-2.5 v2v selector ltx25-v2v supports two extra control modes: outpaint extends/expands the video canvas (e.g. make a vertical clip widescreen, or add space in a direction) — it is positional and mask-free, anchored with a position (center|top|bottom|left|right) and an optional target aspect ratio (16:9|9:16|1:1|4:3|3:4|21:9), and the canvas only grows, never crops; inpaint regenerates a masked region of the source video and requires a mask image (white pixels = region to regenerate) in direct CLI/SDK mode. The hosted video_to_video tool selects these with controlMode outpaint/inpaint and can derive an inpaint mask when the user did not upload one. The direct CLI and sogni-client SDK example expose them via --control-type / control-type (canny|pose|depth|detailer|outpaint|inpaint), with --outpaint-position for outpaint and --mask for inpaint. Pin ltx23-v2v only for rollback. See references/video-editing.md for details.

---

LTX-2.x Prompting Guide

When you use LTX-2.5 or an LTX-2.3 rollback model, do not feed it short tag prompts like "cinematic drone shot over tropical cliffs". LTX renders more reliably from a dense natural-language scene description.

• Write one unbroken paragraph — no line breaks, bullets, headers, or tag blocks.
• Use 4–8 flowing present-tense sentences describing one continuous shot, not a montage.
• Start with shot scale and scene identity, then cover environment, time of day, textures, and named light sources.
• Keep characters and objects concrete and stable; describe one main action thread from start to finish.
• For dialogue, include the exact spoken words in double quotes with the speaker and delivery identified inline.
• Express mood through visible behavior, motion, and sound cues — not vague adjectives.
• Use positive phrasing. Avoid script formatting, negative prompts, on-screen text/logo requests, and filler words like “beautiful” or “nice”.
• Match scene density to clip length. For short clips, describe one main beat, not several actions.

Example rewrite:

User ask: "make a 4k video of a woman in a neon alley"

LTX prompt: "A medium cinematic shot frames a woman in her 30s standing in a rain-soaked neon alley at night, violet and amber signs reflecting across the wet pavement while warm steam drifts from street vents. She wears a dark trench coat with damp strands of black hair clinging near her cheek as light glances across the fabric texture and the brick walls behind her. She turns toward the camera and steps forward with measured focus, one hand tightening around the strap of her bag while rain taps softly on the metal fire escape and a distant train hum rolls through the block. The camera performs a slow push-in as her jaw sets and her breathing steadies, maintaining smooth stabilized motion and a tense urban-thriller mood."

---

Photobooth (Face Transfer)

Generate new stylized portraits from a face photo using InstantID ControlNet:

sogni-agent --photobooth --ref face.jpg "80s fashion portrait"
sogni-agent --photobooth --ref face.jpg -n 4 "LinkedIn professional headshot"

Uses SDXL Turbo (coreml-sogniXLturbo_alpha1_ad) at 1024×1024 by default. The face image is passed via --ref and styled by the prompt. Cannot be combined with --video or -c / --context.

--photobooth is face-reference generation, not full-image editing. If the request is “same image, different style” — for example an anime version that must keep the same face, pose, clothing, background, framing, and composition — use Qwen image editing with -c/--context instead.

Multi-angle mode (--multi-angle / --angles-360) auto-builds the `` prompt and applies the multiple_angles LoRA. --angles-360-video generates i2v clips between consecutive angles (including last → first) and concatenates them with ffmpeg into a seamless loop.

--balance / --balances does not require a prompt and prints the account username, subscription plan (Plan: Sogni Unlimited (active) / Plan: none), and current SPARK and SOGNI balances before exiting. In --json mode the payload carries username and subscription fields (null when unavailable). --doctor likewise reports the authenticated user and an explicit plan check, so a wrong-account API key or a missing subscription is visible at a glance.

---

Personas, Memory, and Personality

Personas

Named people with saved reference photos and optional voice clips for identity-preserving generation:

# Add a persona
sogni-agent --persona-add "Mark" --ref face.jpg --relationship self --description "30s male, brown hair"

# Add with voice clip for video voice cloning
sogni-agent --persona-add "Sarah" --ref sarah.jpg --relationship partner --voice-clip voice.webm

# Generate using a persona (auto-injects photo as context)
sogni-agent --persona "Mark" -o hero.png "superhero in dramatic lighting"

# Video using a persona photo + saved voice identity
sogni-agent --video --persona "Sarah" 'SARAH: "This is my voice."'

# List / remove
sogni-agent --persona-list
sogni-agent --persona-remove "Mark"

Stored at ~/.config/sogni/personas/. Personas resolve by explicit saved name, id, or tag/alias; relationship phrases are not treated as persona identifiers.

Memory (persistent preferences)

Save preferences that agents respect across sessions:

sogni-agent --memory-set preferred_style "watercolor and soft lighting"
sogni-agent --memory-set aspect_ratio "16:9"
sogni-agent --memory-list
sogni-agent --memory-remove preferred_style

Stored at ~/.config/sogni/memories.json.

Personality (custom agent instructions)

Tell the agent how it should behave:

sogni-agent --personality-set "Be concise, always use cinematic lighting"
sogni-agent --personality-get
sogni-agent --personality-clear

Stored at ~/.config/sogni/personality.txt.

---

Hosted API Modes

Hosted API modes require SOGNI_API_KEY.

Choosing a mode. Whatever is driving this CLI is usually a more capable planner than Sogni’s hosted model, so prefer to plan yourself and let the server execute: direct-to-SDK flags for one-shot work, and --api-workflow with an explicit --workflow-input step graph for multi-step/durable work (you author the plan; the server runs it durably with replay — no hosted re-planning). Use --api-chat / --durable-chat when you deliberately want the hosted model to own a long server-side loop, or when several local files must be uploaded for one turn.

• --api-chat targets /v1/chat/completions with Sogni creative-agent tools and delegates planning/tool-selection to the hosted model — reach for it when the caller is a thin client, when you want the hosted model to drive a long server-side tool loop, or when several local files must be uploaded for one turn. The CLI sanitizes prompt-injection markers before forwarding messages and can use the current server-side creative-agent media tools, including video extension, segment replacement, overlays, subtitles, stitch/orbit/dance composition, and generated artifact indexing. Tune with --api-tools creative-agent|creative-tools|none, --no-api-tool-execution, --llm-model, and --system.
• Sogni Intelligence controls include --task-profile general|coding|reasoning, --max-tokens, and --thinking / --no-thinking, which forward to /v1/chat/completions as task_profile, max_tokens, and chat_template_kwargs.enable_thinking. Use --list-api-models or --get-api-model to inspect /v1/models.
• --durable-chat starts a hosted /v1/chat/runs record through the SDK transport. Set SOGNI_SKILL_USE_SDK_TRANSPORT=1 before using it. The CLI streams assistant deltas and de-duplicated per-job progress / ETA / result lines from hosted run events.
• --api-workflow targets /v1/creative-agent/workflows for durable, async workflow records with event streaming and cancellation. Requests carry input.steps plus snake_case controls such as token_type, media_references, max_estimated_capacity_units, and confirm_cost.
• --workflow-input forwards exact durable workflow JSON ({ title?, steps: [...] }). Use this when you need exact multi-step behavior such as repeated replace_video_segment steps with replacementStartSeconds / replacementEndSeconds for interleaved video slices.
• --api-workflow storyboard-video generates a storyline, creates a single storyboard sheet with GPT Image 2.5 Sunburst, then passes that artifact into Seedance 2.5 for 1080p video generation. The -Q fast|hq|pro preset maps to GPT Image low/medium/high quality for that storyboard sheet; explicit model and resolution choices still win.
• Media references from -c, --ref, --ref-end, --keyframe, --ref-audio, --reference-audio-identity, and --ref-video are forwarded as media_references metadata in hosted API requests; --api-chat / --durable-chat also tell the hosted agent which upload to pin at which time as MiniMax H3 keyframes. API chat also attaches image refs as vision inputs. Local file references are uploaded to Sogni media storage first, then forwarded as retrievable URLs so durable executors do not depend on data: URI support. Durable workflow JSON can bind those references into step arguments with sourceStepId: "$input_media". Use direct CLI mode for private media that must not leave the local machine.
• Cost controls use --workflow-max-cost to reject workflow starts above a capacity-unit ceiling, and --confirm-cost / --no-confirm-cost to forward explicit billing confirmation.
• Manage runs with --watch-workflow, --workflow-events, --stream-workflow, --list-workflows, --get-workflow, --cancel-workflow, and --resume-workflow. Use --workflow-input to provide exact durable workflow JSON.
• Replay records use /v1/replay/records: --list-replays [limit], --get-replay , and --ingest-replay expose redacted RunRecord storage for Sogni Intelligence replay/debug viewers.

Override the API origin with --api-base-url, SOGNI_API_BASE_URL, or SOGNI_REST_ENDPOINT. Hosted API credentials are only sent to https://api.sogni.ai by default. Add trusted custom hosts with SOGNI_API_ALLOWED_HOSTS; loopback or non-HTTPS local testing requires SOGNI_ALLOW_UNSAFE_API_BASE_URL=1.

---

Dynamic Prompt Variations

Generate diverse images in a single call with {option1|option2|option3} syntax:

# 3 images: "a red car", "a blue car", "a green car"
sogni-agent -n 3 "a {red|blue|green} car"

# Multiple groups cycle independently
sogni-agent -n 4 "a {cat|dog} in a {garden|kitchen}"
# -> "a cat in a garden", "a dog in a kitchen", "a cat in a garden", "a dog in a kitchen"

Options cycle sequentially per image. Without {...} syntax, -n produces multiple images with the same prompt.

For video, use the same pattern when every output shares the same source/end assets and settings and only the prompt text varies:

sogni-agent --video --ref hero.png -n 3 --duration 5 \
  "{the subject smiles and waves|the subject turns toward the window|the subject raises a hand in greeting}"

If each clip needs different source images, end frames, durations, audio slices, or other per-output settings, keep those as separate per-clip workflow arguments instead of collapsing them into a Dynamic Prompt branch.

---

Token Auto-Fallback

Use --token-type auto to retry native Sogni models with SOGNI tokens when SPARK is insufficient:

sogni-agent --token-type auto "a dragon eating tacos"

Tries SPARK first, then falls back to SOGNI if the balance is too low. Vendor models such as GPT Image 2, Seedance, HappyHorse, and Wan 3 require Premium Spark eligibility and never use SOGNI fallback. If usable balance is still insufficient, buy Spark Packs at https://docs.sogni.ai/pricing/#spark-packs.

On a Sogni Unlimited subscription, Sogni-hosted generation is covered by the plan instead of spending tokens — see the next section.

---

Sogni Unlimited Subscription

Sogni Unlimited is a flat-rate subscription that covers Sogni-hosted (Supernet) image, video, and music generation under a fair-use policy, instead of spending Spark or SOGNI per render. Manage subscriptions where they were purchased — the Stripe billing portal for web checkouts, or the App Store / Google Play account settings for mobile.

Plans

App Store and Google Play prices may differ from web pricing due to platform fees. A 3-day free trial is available once per account (a payment method is required and the subscription converts to paid when the trial ends unless cancelled first).

Plan pricing, included features and models, usage allowances, fair-use controls, and other limits are subject to change at Sogni AI’s discretion, subject to applicable law. Sogni will provide advance notice of material changes affecting an active paid subscription when required. Treat the live plan catalog and checkout as authoritative.

What the subscription covers

• Covered: Sogni-hosted models on the Supernet — image, video, and music generation, including worker-hosted premium models. Covered renders bill to the subscription and do not spend Spark or SOGNI.
• Not subscription-covered: external-vendor models — GPT Image 2 / GPT Image 2.5 Sunburst / GPT Image 2.5 Flare (gpt-image-2, gpt-image-2.5-sunburst, gpt-image-2.5-flare), Seedance 2.0 / Seedance 2.0 Mini / Seedance 2.0 Mini Uncensored / Seedance 2.0 Fast / Seedance 2.5 / Seedance 2.5 Uncensored (seedance-2-0, seedance-2-0-mini, seedance-2-0-mini-uncensored, seedance-2-0-fast, seedance-2-5, seedance-2-5-uncensored), HappyHorse 1.1 (happyhorse-1.1-t2v, happyhorse-1.1-i2v, happyhorse-1.1-r2v), Wan 3 (wan3.0-video), and Wan 3.0 Enhanced (wan3.0-spicy-video). They never fall back to SOGNI. Wan 3.0 Enhanced may use its separate launch-credit ledger before PAYG; that credit does not make the model subscription-covered.
• Token choice stays yours: selecting SOGNI (--token-type sogni) opts a job out of subscription coverage and spends SOGNI instead. Coverage applies when the active token is Spark.

By default the CLI sends no billingMode/coverage hint; the server decides coverage from the account’s verified entitlement and the resolved model, and a subscription claim is never honored without a server-verified entitlement. --billing-mode makes the choice explicit when you need it: subscription requires Unlimited coverage (the job fails instead of spending tokens), tokens opts out of coverage and bills Spark/SOGNI, and auto states the default server behavior explicitly.

Do not use tokenType: "spark" by itself to determine that Spark paid for a render. tokenType is also the quote/accounting denomination for covered jobs. The server’s separate paymentModel is authoritative: subscription means it skipped the artist Spark/SOGNI debit, while paid_spark, free_spark, and sogni identify token-funded paths. Some client result summaries do not expose paymentModel; in that case the payment source is unknown from that result alone. A request that completes successfully with --billing-mode subscription was covered—if coverage is unavailable, the server returns 4078 or 4080 instead of silently spending Spark.

With an active subscription, the CLI also skips its client-side “insufficient SPARK” pre-flight for covered video renders — a low token balance no longer blocks jobs the plan pays for. Vendor models and --billing-mode tokens keep the pre-flight, and the server remains authoritative either way.

Free-trial access

Trials include evaluation limits on generation volume, media size, and API access. Full plan limits apply once the trial converts to paid. Cancelling during the trial prevents the first charge and keeps trial access, with the same trial limits, until the original trial end date.

Fair-use scheduling

Unlimited is fair-use, not unmetered. Published plan limits are:

• Unlimited: up to 4 concurrent image jobs and 1 concurrent video job, or 2 video jobs with fast/turbo models; queue up to 64 media, including up to 8 videos.
• Unlimited Pro: up to 16 concurrent image jobs and 4 concurrent video jobs, with standard MiniMax H3 limited to 2; queue up to 192 media, including up to 24 videos.

Actual throughput varies with demand, available Supernet capacity, and fair-use controls. Sustained high-volume or automated use may be delayed or rate-limited. Unlimited Pro receives higher subscription queue priority than Unlimited, and Premium Spark remains available for fastest paid priority. Subscription jobs cannot target specific workers.

Billing states & cancellation

• Active / trialing: covered renders run normally.
• Cancellation (paid): Unlimited access continues until the end of the period already paid for; it simply does not renew.
• Cancellation (during trial): no charge is made, and trial access continues until the original trial end date. Cancelling does not extend the trial or make the account eligible for another one.
• Grace / payment retry: if a renewal payment fails, the provider retries it and Unlimited access is paused during the retry window — covered renders are declined with a renewal-retry error, and access resumes automatically once payment succeeds. You can keep rendering with Spark or SOGNI in the meantime.
• Refunds: mid-term refunds are not offered by default; App Store / Google Play purchases follow the store’s refund process, and Stripe (web) refunds are handled by Sogni support.

Subscription billing errors

When a generation cannot bill to the subscription, the CLI surfaces a structured error (--json includes errorCode, errorCategory, and a hint):

Worker revenue share

Sogni workers that power subscription-covered jobs earn from a separate monthly pool — 51% of net subscription revenue — settled per UTC month and claimable in USDC on Base. Subscription jobs are excluded from the regular Spark/SOGNI token-economy leaderboard (they do not spend tokens) and accrue to this pool instead.

---

Error Reporting & Output

• Exit codes: failures use a non-zero exit code with human-readable stderr.
• Structured output: add --json when an agent needs machine-parseable success/error data, or --last to inspect the last render. JSON failures include canonical errorType, errorCategory, and retryable fields where the shared runtime can classify the error.
• Subscription billing errors: subscription-billing failures carry errorCode 4078 / 4079 / 4080 / 4081, errorCategory: "subscription_billing", and an actionable hint. See Subscription billing errors for what each means; in particular, do not auto-retry a 4080 (grace / renewal-retry) covered job — pay with Spark or SOGNI instead.
• stdout stays parseable in --json mode: progress lines, SSE workflow frames, and warnings go to stderr; stdout carries exactly one JSON object. --last --json wraps the record in a { "success": true, ... } envelope and exits 1 with errorCode: "NO_LAST_RENDER" when nothing has been rendered yet.
• Output files: use -o to save locally; otherwise the CLI prints a result URL.
• Quiet mode: -q / --quiet suppresses progress output without changing exit semantics.
• Interrupts: Ctrl-C exits with the conventional signal code and cleans up the CLI’s temporary files.

---

For AI Agents

This skill is designed to be loaded into agent runtimes as a first-class capability.

1. Behavior contract — SKILL.md The canonical instructions for how the agent should call sogni-agent. Load this as the skill source.
2. Install/setup hints — llm.txt A condensed install/setup reference for agents that fetch llm.txt over HTTPS: https://raw.githubusercontent.com/Sogni-AI/sogni-creative-agent-skill/main/llm.txt
3. OpenClaw manifest — openclaw.plugin.json Plugin metadata, config schema, and defaults for OpenClaw-aware runtimes.
4. Structured output — --json Use --json for machine-readable success/error payloads. Use --last to read the previous render’s metadata.
5. Agent-safe install/upgrade Prefer the npm install -g and git -C "$DEST" pull --ff-only paths above. Avoid generating clone-or-pull bootstrap scripts with set -e, bash -c, sh -c, or inline repository URLs — agent sandboxes correctly route those through approval and the install will stall.
6. Verify with doctor After any install or upgrade, run sogni-agent doctor --json and confirm "success": true before reporting the install as working.
7. Update notices for agents When a newer version exists, any command may print one advisory stderr line — [sogni-agent] Update available: -> ... — at most once per day (stdout JSON is never touched). Agents should relay it to the user and offer sogni-agent self-update, or run sogni-agent --snooze-update if the user declines. Interactive TTY users get a banner instead. Each failed check carries a detail string with the fix.
8. SSRF / URL safety The CLI validates every HTTP(S) media reference with an SSRF guard (ssrf-guard.mjs) and re-validates each redirect hop on download. Localhost and private-network URLs are rejected; only public HTTPS references are forwarded as Seedance multimodal context.

---

Development

Run the unit test suite (works without any Sogni credentials or private repos):

npm test

Paid integration tests are opt-in: npm run test:integration (requires a Sogni API key and submits real GPU jobs).

Architecture notes, the private-runtime sync workflow, code-placement policy, and the release process live in CONTRIBUTING.md.

Issues and feature requests: github.com/Sogni-AI/sogni-creative-agent-skill/issues.

---

License

MIT © Sogni AI

Personal LoRA imports and library management: read Personal LoRAs for account-bound discovery, asynchronous import status, consent, compatible models, and rendering with owned adapters.