Tenzo

A local-first studio for designing, refining, and packaging AI character cards.

Tenzo turns character creation into a focused writing workflow. Build a character by hand, develop one with an AI assistant, tune each field with reusable instructions, and export a portable card for SillyTavern and other V2-compatible tools.

No account or hosted database is required. Your character library, portraits, templates, and conversations live in your browser.

Highlights

	Feature	What it gives you
✨	Character Assistant	Explore concepts, fill gaps, and revise one field or a whole card through conversation. AI edits arrive as reviewable proposals, so you stay in control.
📝	Purpose-built editor	Author core identity, dialogue, alternate greetings, character books, prompt overrides, metadata, tags, and custom fields in one workspace.
🪄	Field-level generation	Generate, continue, or rewrite individual fields with streaming output, sampling controls, and token estimates.
📑	Reusable templates	Save field instructions, scope them to specific fields, and use example characters as style and structure references.
🖼️	Portrait studio	Upload PNG, JPEG, WebP, or GIF artwork and adjust its focal point for a SillyTavern-ready 2:3 portrait.
🗃️	Local character library	Keep multiple works in progress with automatic browser persistence and switch between them instantly.
📦	Portable import and export	Read V1/V2 JSON and PNG cards, preserve unknown extension data, and export hybrid V1+V2 JSON or PNG cards with embedded metadata.
💾	Bulk export and backup	Package selected characters as ZIP or tar.gz, or back up the full workspace without including API credentials.

See Tenzo in action

How it fits together

flowchart LR
    Idea[Idea or existing card] --> Studio[Tenzo workspace]
    Examples[Example characters] --> Studio
    Portrait[Portrait artwork] --> Studio

    Studio <--> Library[(Browser storage)]
    Studio --> Assistant[AI assistant]
    Assistant --> Review{Review proposed edits}
    Review -->|Apply| Studio
    Review -->|Reject| Studio

    Studio --> JSON[Hybrid V1 + V2 JSON]
    Studio --> PNG[PNG with embedded card data]
    Studio --> Archive[ZIP / tar.gz archive]

Loading

The assistant can focus on a single field or reason about the complete character. It streams suggestions into a separate review step; conflicting edits are detected before changes are applied to the card.

Local-first by design

flowchart TB
    Browser[Your browser]
    Storage[(IndexedDB + local settings)]
    Route[Stateless Tenzo AI route]
    Provider[Your configured AI provider]

    Browser <--> Storage
    Browser -->|Generation request + API credential| Route
    Route -->|OpenAI-compatible request| Provider
    Provider -->|Streamed response| Route
    Route --> Browser

Loading

• Character data, images, assistant sessions, and templates are persisted locally.
• Card parsing, portrait processing, PNG metadata embedding, archives, and backups run in the browser.
• AI credentials are supplied per request and are never included in workspace backups or persisted by the server route.
• Tenzo works with OpenAI-compatible endpoints, including model discovery and capability checks where the provider supports them.

Important

Browser storage is tied to the current browser profile and origin. Export a full workspace backup before clearing site data or moving to another device.

Under the hood

Tenzo keeps the interesting machinery close to the browser instead of hiding it behind a conventional application backend.

• Binary assets stay binary. Portraits are stored as Blob objects in IndexedDB rather than inflated base64 strings. A bounded, shared LRU cache retains object URLs for recently viewed portraits, deduplicates concurrent loads, and revokes URLs on eviction. That makes character switching feel immediate without leaking browser memory.
• PNG cards are edited without recompressing artwork. Tenzo reads the PNG chunk stream, removes stale chara or ccv3 metadata, inserts a UTF-8/base64 chara tEXt chunk immediately before IEND, and reassembles the original chunks. The image data itself is never decoded and re-encoded.
• Local data has application-level migrations. Dexie provides stable IndexedDB stores while Tenzo tracks record migrations separately. A destructive migration must declare a warning, lock the editor, and require the user to download a recovery backup before it can run.
• The UI writes optimistically. Persistent collections update React immediately and roll back the local view if IndexedDB rejects the mutation, combining local-app responsiveness with durable storage semantics.
• Assistant changes are transactions, not pasted prose. The model produces typed patches for fields, tags, greetings, custom fields, and character books. Tenzo can detect overlapping edits and stale revisions, then lets the user apply or reject proposals explicitly.
• Generation is streamed and cancellation-aware. Stateless TanStack routes bridge OpenAI-compatible providers through server-sent events. Browser cancellation propagates upstream, while provider failures are normalized into actionable stream errors.
• Backups are real archives. Full-workspace ZIP and tar.gz files contain a versioned manifest, character records, examples, settings, and original portrait blobs. API credentials are deliberately blanked before packaging.

Card workflow

1. Create a character or import an existing .json or .png card.
2. Write directly, generate individual fields, or collaborate with the Character Assistant.
3. Add example cards and reusable templates when you want consistent style or structure.
4. Upload and frame a portrait, then review token estimates and card metadata.
5. Export the current card, a selected collection, or a complete workspace backup.

Tenzo preserves unknown extensions values during import and export, including character-book and entry-level extension data. Macros such as {{char}} and {{user}} remain intact.

Run locally

Requirements

• Node.js 24
• pnpm 10 or newer

git clone https://github.com/CatOfJupit3r/tenzo.git
cd tenzo
pnpm install
pnpm run dev

Open http://localhost:3030, then configure an OpenAI-compatible endpoint, model, and API key in Settings -> Connection to enable AI features. The editor and local card tools do not require an AI connection.

Development

Tenzo is a TypeScript monorepo centered on a React 19 application in apps/web, built with TanStack Start and Router, TanStack AI, Tailwind CSS 4, TipTap, Dexie, Jotai, Zod, Vitest, and Playwright.

Command	Purpose
pnpm run dev	Start the development app
pnpm run build	Create a production build
pnpm run test	Run the test suite
pnpm run check-types	Check TypeScript types
pnpm run lint	Run ESLint
pnpm run prettify	Check formatting

The character-card implementation lives in apps/web/src/features/character-creator. Product scope and implementation status are tracked in the character card creator roadmap.

License

Tenzo is available under the MIT License.