Drop a seed. Walk a universe.
The stories you keep imagining. See them, hear them, share them.
You don't ask for a story. You drop a seed and walk every "what if" branch the universe could take. Every splice becomes a real branch: the agent re-harmonises the canon, paints the visuals, clones a fitting narrator voice, and renders the version you picked into a video you can post.
| Action | Result |
|---|---|
| Type a seed | A canon-tracked, live branching story canvas |
| Pick an option | Hermes proposes the next 2 to 3 critical beats |
| Splice a beat mid-tree | Downstream beats re-harmonise themselves automatically |
| Hit Render | A landscape audiobook or a vertical reel, ready to post |
| Drop a YouTube URL | ElevenLabs IVC clones that voice for every line of narration |
| Save / Load | Full tree state, with provenance, in localStorage |
Two distinct loops sit under the canvas. The authoring loop runs per beat as you grow the tree. The render loop runs once, when you press Render: the agent dispatches four registered Hermes tools in sequence on a single chat completion, no app-side orchestration.
The canvas exposes two + buttons with different shapes and different
jobs. Both grow the tree, but only one of them needs the council.
flowchart LR
classDef canon fill:#0d1828,stroke:#4fc3f7,color:#e0f7fa,stroke-width:2px
classDef inserted fill:#1a1208,stroke:#e9c16b,color:#fff8e1,stroke-width:2px
classDef whatif fill:#0f1019,stroke:#7a7e8a,color:#cdd2de,stroke-dasharray: 5 3
classDef arm fill:#0f131b,stroke:#e9c16b,color:#e9c16b,stroke-width:2px
classDef trunk fill:#0f131b,stroke:#7a7e8a,color:#cdd2de,stroke-width:2px,stroke-dasharray: 4 3
A["Beat 1<br/>opening"]:::canon
B["Beat 2<br/>discovery"]:::canon
C["Beat 3<br/>climax"]:::canon
ARM["⬛ +<br/>arm splice"]:::arm
INS["Beat 1.5<br/>inserted, gold"]:::inserted
TRK(("○ +<br/>trunk")):::trunk
WIF["Beat 2'<br/>what-if branch"]:::whatif
A -- canon --> ARM
ARM -- splice --> INS
INS -- canon --> B
B -- canon --> C
A -. parent stub .-> TRK
TRK -. what-if .-> WIF
| Gesture | Where it sits | What it does |
|---|---|---|
| Square + (gold, parallelogram) | Midpoint between a parent and its canon child | Splices a new beat into the canon. narrative-writer-assist drafts the inserted beat, then the council cascade patches every downstream beat the splice made stale. |
| Round + (dashed circle) | Off the parent's outgoing stub | Adds a what-if sibling, a parallel timeline that doesn't touch existing canon. No cascade fires; brainstorm just produces 3 fresh next options on this new branch. |
Every beat passes through a small council of narrative skills. Splice one beat and the cascade walks downstream by itself until the canon is coherent again.
sequenceDiagram
autonumber
participant U as user
participant N as next.js
participant H as hermes gateway
U->>N: enter seed
N->>H: brainstorm
H-->>N: 2-3 next options
U->>N: pick / splice
rect rgba(80,140,255,0.10)
Note over N,H: re-harmonisation cascade
N->>H: continuity-critic
H-->>N: contradictions per node
N->>H: rewriter
H-->>N: patched beats
N->>H: judge
H-->>N: approve / reject
end
N-->>U: tree updated
When you press Render, the agent chains four registered Hermes tools in sequence on a single chat completion. No app-side glue.
sequenceDiagram
autonumber
participant A as agent
participant D as narration_voice_director
participant S as voice_sample_builder
participant C as voice_clone_synthesize
participant R as noustiny_storybook
A->>D: title, seed, voice_gender
D-->>A: persona_label, search_query, fallback_query
A->>S: query
S-->>A: sample_path (24 kHz mono PCM)
A->>C: text, reference_wav
C-->>A: audio + per-character alignment
A->>R: voice_reference_wav, persona, render args
R-->>A: rendered video path
Stories drift visually when "Tony Stark", "Mr. Stark", and "Tony" each get a different portrait. Noustiny pins the cast at seed time and threads the same reference image through every beat that mentions them.
flowchart LR
classDef seed fill:#0d1828,stroke:#4fc3f7,color:#e0f7fa,stroke-width:2px
classDef sheet fill:#1a1208,stroke:#e9c16b,color:#fff8e1,stroke-width:2px
classDef beat fill:#0f131b,stroke:#7a7e8a,color:#cdd2de
classDef tool fill:#0a1018,stroke:#4fc3f7,color:#4fc3f7,stroke-dasharray: 4 3
seed["seed prompt"]:::seed --> sheet
sheet["character_sheet_builder<br/>1 to 4 named cast<br/>IP-free portrait per character"]:::sheet
beat["beat mentions a name"]:::beat --> reg
reg{{"character_alias_resolver<br/>'Mr. Stark' → 'Tony Stark'"}}:::tool --> lookup
lookup{{"character_registry_lookup<br/>fetch portrait reference"}}:::tool --> img
sheet -. registers .-> lookup
img["beat image gen<br/>conditioned on portrait"]:::beat
The cast sheet is built once. From then on, every beat that names a character pulls the same portrait reference, so a face stays a face across the whole tree, including the rendered video.
12 registered tool names across 10 Python files. Each one fills a primitive Hermes did not ship with.
| Tool | Purpose |
|---|---|
story_tree_graph |
Tree graph operations: canon path, descendants, splice insertion points |
narrative_context_builder |
Walks the canon chain, returns the live structured context every skill reads from |
motif_tracker |
Recurring motif memory across the arc (a sword in beat 2 must reappear later) |
| Tool | Purpose |
|---|---|
story_copyright_detector |
IP scrub: "Iron Man" becomes an IP-free description before the image API sees it |
character_sheet_builder |
Produces 1 to 4 named characters with IP-free visuals plus portrait prompts |
character_registry_lookup |
Find a character on the cast sheet, attach the correct portrait reference |
character_alias_resolver |
"Mr. Stark" resolves to "Tony Stark" so one character keeps one portrait |
| Tool | Purpose |
|---|---|
narration_voice_director |
Director-agent reads seed and story, returns persona plus reference search query |
voice_sample_builder |
yt-dlp + ffmpeg, accepts URL, 11-char ID, or free text. Normalises to 24 kHz mono |
voice_clone_synthesize |
ElevenLabs IVC + with-timestamps. Voice cached by reference SHA. Alignment is free |
voice_clone_cleanup |
Frees the cached voice ID after render |
| Tool | Purpose |
|---|---|
noustiny_storybook |
Render entry. Agent dispatches it; one call drives the FastAPI service end to end |
13 agentskills.io bundles, each one a system prompt for a single creative concern. Next reads them from the local mirror and sends them as the system role on every chat completion; the autonomous storybook chain pulls its own from the gateway side.
| Skill | Role |
|---|---|
narrative-brainstorm |
Proposes 2 to 3 next-checkpoint options from the canon chain |
narrative-writer-assist |
Writes a spliced insert beat that fits parent and child |
narrative-continuity-critic |
Audits downstream beats against any insert |
narrative-rewriter |
Patches the stale beats the critic flagged |
narrative-judge |
Approves or rejects the rewrite against the original |
narrative-scene-qa |
Per-beat sanity check (consistency, length, register) |
narrative-writer |
Seals the chosen branch as final prose |
| Skill | Role |
|---|---|
story-copyright-detector |
Skill counterpart of the same-named tool |
character-sheet-builder |
Cast extraction rules + IP-free portrait prompt format |
visual-prompt-builder |
Turns a beat into an IP-free image prompt; reads character sheet |
scene-composition |
Shot framing and layout rules |
storybook-intro |
Cinematic intro page generator |
| Skill | Role |
|---|---|
narration-voice-director |
Persona reasoning rules backing the same-named tool |
flowchart TB
subgraph windows[Windows side]
nx[Next.js<br/>:3000]
mirror[hermes-skills-local<br/>NTFS mirror]
end
subgraph wslblock[WSL]
gw[Hermes gateway<br/>:8642]
svc[Storybook render<br/>:8643]
tools[hermes-agent tools]
skills[hermes-agent skills<br/>canonical]
end
subgraph cloud[Cloud]
nous[Nous Portal]
el[ElevenLabs]
yt[YouTube / yt-dlp]
end
browser([browser]) --> nx
nx -- chat completion --> gw
nx -. "read SKILL.md" .-> mirror
gw -- forward --> nous
gw --> tools
tools --> el
tools --> yt
tools --> svc
svc --> nx
The hermes-skills-local directory is a Windows-side mirror of the canonical
skills folder in WSL. Reading SKILL.md over UNC stalls under parallel reads
(25 second timeouts on canvas-enter), so Next reads from local NTFS instead.
Resync after editing a SKILL.md:
cp -r //wsl.localhost/Ubuntu/home/<user>/hermes-agent/skills/creative \
./hermes-skills-local/.
├── hermes-additions/ tools + skills we authored for Hermes
│ ├── tools/ 10 .py files, 12 registered tool names
│ ├── skills/creative/ 13 SKILL.md bundles
│ └── README.md install + integration notes
├── hermes-skills-local/ Windows mirror (gitignored)
└── web/ Next.js 16 app
├── app/ routes, pages, layouts
├── components/ canvas, modals, agent ticker, etc.
├── lib/ store, hermes client, save system
└── public/ static assets
Nous Creative Hackathon, 2026.
Hermes Agent · Nous Research ElevenLabs Instant Voice Cloning yt-dlp ffmpeg + libass