Kumanga is a free, open-source (MIT) AI manga creator that runs in your own browser. AI generates reusable assets — characters, poses, expressions, backgrounds — a non-destructive editor composes them into manga pages, and a natural-language Manga Agent can build whole panels for you. Bring your own API key (BYOK); there is no account, no cloud lock-in, and no bundled model.
Kuma (bear) + manga. Created and maintained by BotTony329.
🌐 Website · 🎬 Try the live studio · 📚 Studio docs · 🏛 Architecture
Kumanga is an AI manga harness built on one thesis: AI creates reusable manga assets; the creator composes those assets into manga. Generation and composition are separate systems, and you are always the director. Think Figma/Canva + manga editor + AI character studio + AI asset library — plus a Manga Agent that operates the same editor through natural-language prompts.
- Stack: Next.js 15 · React 19 · TypeScript 5 · npm workspaces
- Status: v0.1 open-source preview, release-frozen 2026-08-22 —
991/991 tests, 98/98 security tests,
npm auditclean (freeze report) - License: MIT (keep
NOTICE.mdattribution on forks)
Because one-off images don't make a manga. A 20-page chapter needs the same character, on-model, from panel to panel — plus framing, staging and dialogue you control. Kumanga treats AI output as an asset library, not a slot machine:
| Need | Generic image AI | Kumanga |
|---|---|---|
| Same character every panel | Re-prompt and pray | Reusable character assets — structured poses × expressions collections, reference-image-aware generation |
| Full shot → close-up | New generation each time | Crop modes: one full-body asset becomes a full, medium or close-up shot with zero new generations |
| Framing & staging | Not supported | Non-destructive panel editor (Figma-frame semantics); instances never modify their source asset |
| Whole-page help | — | Manga Agent: natural language → plan → validated tool calls → an editable page |
| Where it runs | Vendor cloud | Your machine (localhost out of the box); optional self-deploy |
| Cost & lock-in | Subscription | MIT, free, your own keys |
- Asset library, not one-off images. Characters are structured collections (poses × expressions), browsable visually. Backgrounds, props and uploads are first-class reusable assets.
- Non-destructive panel editor. A panel is a clipping viewport; crop modes re-frame assets without new generations; instances never modify their source.
- Real AI generation. A provider abstraction with a Google Gemini adapter (reference-image aware, for character-consistent poses and expressions) and a generic OpenAI-compatible REST adapter.
- Manga Agent. Prompt → skill-guided plan → validated tool calls → execution through the same editor commands the manual UI uses.
- Persistence & export. Projects survive refresh (IndexedDB + remote object storage for images); pages export to PNG at 1× / 2×.
You describe the page in plain language; the agent returns a finished, fully-editable page region. The LLM acts as a creative director while code alone owns state, identity and execution: prompts pass through a literal lock, one server-side planning call, structural and semantic validation, project resolution (names bound to real IDs), and a transactional executor with rollback. Because agent runs go through ordinary editor commands, one Undo reverts a whole agent run and every result stays hand-editable. See ARCHITECTURE.md for the full pipeline.
Kumanga never ships a model key. Open AI Settings inside the app, pick a standard, enter base URL + key + model, Test Connection, Save:
| Surface | Supported standards |
|---|---|
| Image generation | Google Gemini (reference-image aware) · any OpenAI-compatible REST endpoint |
| Manga Agent | OpenAI-compatible · Anthropic-compatible · Gemini |
Credentials are AES-GCM-encrypted into HttpOnly session cookies — never client-readable, never stored in project data. Replace or forget them any time, no redeploys.
Yes — Kumanga is local-first. Projects, assets and generated images live in
your browser storage (IndexedDB, plus object storage for images when you deploy
it), not in our cloud. There is no account system. Everything works on
localhost out of the box; deploying to Vercel is optional, never required.
The deployment surface is audited: SSRF/redirect/size caps, object-storage
traversal and provider-session security all have dedicated test suites
(98/98 passing at the v0.1 freeze).
Requires Node.js 18.18+ (20+ LTS recommended for Next.js 15).
git clone https://github.com/BotTony329/mangaharness.git
cd mangaharness
npm install
cp .env.example .env.local # optional — the editor works without any AI key
npm run dev # http://localhost:3000No registration, no account, no cloud setup. To use AI, add your own key in AI Settings inside the app.
| Command | What it does |
|---|---|
npm run dev |
Start the studio at localhost |
npm run build / npm start |
Production build / serve it |
npm test |
Full test suite (domain, geometry, security, agent) |
npm run typecheck / npm run lint |
Static gates |
Yes. Kumanga is open source under the MIT license — clone it, run it, fork it.
You supply your own AI API keys (BYOK), so there is no Kumanga subscription and
no bundled model. If you fork or reuse it, keep the attribution in
NOTICE.md intact.
No account, ever. The editor itself works with no AI key at all — you can compose pages and explore the studio first. When you want AI generation or the Manga Agent, open AI Settings and connect your own provider key.
Projects and assets live in your browser's IndexedDB (with object storage for images on deployments); they survive refresh and never leave your environment unless you deploy your own instance. API keys are AES-GCM-encrypted into HttpOnly session cookies — they are never client-readable and never written into project data, so exports and shares don't leak credentials.
That is the core design. Characters are structured collections of poses × expressions, and the Gemini adapter is reference-image aware: it generates new poses and expressions from your existing character assets instead of re-rolling from scratch. Framing changes (full / medium / close-up) don't require any generation at all — crop modes re-frame the same asset.
Undo once. Agent runs execute through the same editor commands as the manual UI, inside a transaction — a single Undo reverts the entire run, and everything the agent produced remains fully hand-editable afterwards.
Pages export to PNG at 1× or 2× resolution. Project data stays in your local storage, so you keep everything — projects, assets and generated images — even if you stop using the app.
Yes, deploying (e.g. to Vercel) is optional and fully self-serve. Set
APP_ENCRYPTION_KEY (encrypts users' BYOK credentials — it contains no AI key
itself) and optionally BLOB_READ_WRITE_TOKEN; users' own BYOK settings
override any operator-default providers. A reference deployment runs at
mangaharness.vercel.app.
v0.1 open-source preview, release-frozen on 2026-08-22 with a published
baseline: 991/991 tests passing, 98/98 security tests, clean typecheck, lint
and build, and zero npm audit vulnerabilities. Known-good baseline and
architecture notes: docs/RELEASE_FREEZE_V0.1.md
and ARCHITECTURE.md.
manga-studio/— the Next.js app (editor core, agent, services, API routes). See the studio README for configuration details.ARCHITECTURE.md— module boundaries and the agent pipeline.docs/RELEASE_FREEZE_V0.1.md— the v0.1 known-good baseline.
Kumanga is MIT-licensed and built by BotTony329. If you fork or reuse it,
keep the attribution (NOTICE.md) intact.