@skyphusion/create-prism

@skyphusion/create-prism

Scaffold a new prism deployment: a multimodal AI playground (chat across many models, voice chat, image / video / music generation, RAG, and web search) that runs as a single Cloudflare Worker.

This package is the scaffolder. It copies a fresh, deployable prism tree into a new directory; you then wire your own Cloudflare resources and deploy. The full prism source ships inside the tarball, so there is no separate clone step.

Quick start

npm create @skyphusion/prism my-prism
cd my-prism
npm install
npm run bootstrap

npm run bootstrap copies wrangler.example.toml to wrangler.toml (your per-deployer config; gitignored). From there the scaffolder prints numbered next steps; the generated README.md has the complete walkthrough.

You can also run the scaffolder directly:

npx @skyphusion/create-prism my-prism

my-prism is optional (default prism-app). The command refuses to write into a directory that already exists and is not empty; nothing is written in that case.

Requires Node.js 20 or later. Zero runtime dependencies.

What you get

A working single-Worker app for the Cloudflare AI stack: chat over many models and five providers, hands-free voice chat, image / TTS / STT / video / music generation, RAG over uploaded files, projects, Discord ingestion, and opt-in web search. One web UI, per-user history, R2 for all binary artifacts. No framework, no build step beyond TypeScript.

Cloudflare resources you provide

Everything bills through Cloudflare Unified Billing on your own AI Gateway; there is no deployer BYOK key. Create these once and wire their ids into wrangler.toml (the scaffolder prints the exact commands):

Required bindings

Binding	Resource	Purpose
AI	Workers AI + AI Gateway	Every modality funnels through env.AI.run(); paid partners route through your gateway.
DB	D1 database	Chat metadata, conversations, RAG chunk text, projects, accounts.
R2	R2 bucket	All binary artifacts (inputs and generated output). Nothing binary touches D1.
VEC	Vectorize index	RAG embeddings, 768-dim BGE-base, cosine.
LONGRUN	Workflows	Durable execution for long-running video and music generation.

Plus the ASSETS static-assets binding (the public/ frontend) and the STT_SESSION Durable Object for streaming voice, both already declared in the template wrangler.example.toml.

Optional

• SEARXNG_URL (and the SEARXNG_ACCESS_CLIENT_ID / SEARXNG_ACCESS_CLIENT_SECRET halves when your SearXNG instance sits behind Cloudflare Access) enable the opt-in web-search source. Unset, web search silently skips SearXNG; keyless Wikipedia still runs.

Auth: two modes, one identity seam (AUTH_MODE)

The scaffolded worker learns who a caller is one of two ways, chosen by the AUTH_MODE var in wrangler.toml:

• access (default, private self-host). Cloudflare Access sits in front of the worker URL; identity comes from the Cf-Access-Authenticated-User-Email header. Inference bills the deployer, through the worker gateway secrets GATEWAY_ID and CF_AIG_TOKEN.
• public (the open, first-party-signup service). First-party username/password accounts behind an opaque server-side session cookie. Every user brings their own AI Gateway credentials (mandatory BYOK), so the worker holds no gateway secrets and visitor inference never bills the host. Do not set GATEWAY_ID / CF_AIG_TOKEN on a public worker; they are ignored in this mode. No Cloudflare Access needed.

Either way, one stable opaque account id scopes history and R2 ownership, so cross-user access is impossible even if an id is guessed.

License and the AGPL-3.0 network-service obligation

prism (and every app you scaffold from it) is AGPL-3.0-only. If you run it as a network service for users, AGPL-3.0 section 13 requires you to offer those users the Corresponding Source of your running instance under the same license. The template satisfies this with an in-app "Source code" link in the account menu; if you modify the worker and run it publicly, point that link at your own source. This is a legal obligation, not a suggestion: publishing a public prism instance means publishing its source.

Note for offline / mirrored npm installs

The scaffolded app has one dependency that resolves from a URL rather than the npm registry: xlsx (spreadsheet RAG) is pinned to a tarball on cdn.sheetjs.com (https://cdn.sheetjs.com/xlsx-0.20.3/xlsx-0.20.3.tgz), per the SheetJS distribution policy. A normal npm install fetches it fine. If you install behind a private mirror, an air-gapped proxy, or a lockfile-strict CI that blocks non-registry URLs, allowlist cdn.sheetjs.com or vendor that tarball into your mirror. Everything else resolves from the public npm registry.

Links

• prism source and full docs: https://github.com/skyphusion-labs/prism
• Live instance: https://play.skyphusion.org
• Issues: https://github.com/skyphusion-labs/prism/issues

Licensed AGPL-3.0-only. See LICENSE and NOTICE.