DevStack · devstack.bid

🚧 Branch feat/studio · v0.8.0-rc.2 (preview, not deployed yet). This branch merges the Studio workspace into DevStack. devstack.bid still runs v0.7.0 from main, so the /studio/ link below works only after this branch ships. You can run it locally with npm run preview.

🚧 Branch feat/studio · v0.8.0-rc.2 (ยังไม่ deploy) branch นี้รวม Studio เข้ากับ DevStack ตอนนี้ devstack.bid ยังเป็น v0.7.0 จาก main ลิงก์ /studio/ ด้านล่างจะใช้ได้หลัง branch นี้ขึ้นระบบจริง ระหว่างนี้รันในเครื่องได้ด้วย npm run preview

An open developer platform from the NVX community. You can turn setup commands into runnable Stack Templates, chat with free AI models, run an AI agent with a sandbox, bring your own OpenAI key to the Sandbox Factory, or build a full-stack app in your browser with the Studio. The whole site is Thai + English.

แพลตฟอร์มนักพัฒนาแบบเปิดจากคอมมูนิตี้ NVX เปลี่ยนคำสั่งติดตั้งให้เป็น Stack Template ที่รันได้จริง แชตกับ AI ฟรี ให้ agent ทำงานใน sandbox ใช้ OpenAI key ของตัวเองใน Sandbox Factory หรือสร้างแอป full-stack ในเบราว์เซอร์ด้วย Studio ทั้งเว็บใช้ได้ทั้งภาษาไทยและอังกฤษ

Live: https://devstack.bid · Docs / เอกสาร: docs/README.md (also at https://devstack.bid/docs) · AI agents: AGENTS.md, /llms.txt · MIT License

What's inside / มีอะไรบ้าง

App	Path	What it does	AI and cost
Learn	/learn	Launch page for the whole system, made by the creator devstack.sdk@gmail.com and Cursor. The homepage Ask the AI agent button opens it. Every page header has Setup Cursor (create a project or ask Cursor).	—
Stack Builder	/templates, /builder	10 curated Stack Templates (Next.js, Vite, Express, FastAPI, Streamlit, CLI, PyPI …). Pick versions and add-ons, then reorder, share as a link, or download setup.sh / setup.ps1 / a starter zip. The AI planner turns a description into a stack.	Planner: Hugging Face DeepSeek V3, falling back to an offline rule-based planner. You can also bring your own OpenAI key.
Chat	/chat	A free AI chat (General, Thai, Code, Stronger and Image modes) with history saved per browser.	Workers AI (Gemma 4, SEA-LION, Qwen Coder, FLUX) and HF DeepSeek V3, with free daily quotas.
Dev Chat	/chat/dev	The Image and Research Orchestrator agent, with a live sandbox panel (commands, subagents, files).	OpenAI Agents API on our key, capped at a few turns per day.
Sandbox Factory	/factory	Your own agent and sandbox (OpenAI-hosted or self-hosted codex exec-server).	Your OpenAI key, sent per request and never stored.
Studio (beta)	/studio/	Four agents (Core, Code, DB, Design) build a full-stack app in your browser: a static live preview (Sucrase + sandboxed iframe, React only), PGlite + pgvector, AI images, and a one-click export (source, SQL, Prisma, Vercel/Cloudflare/Docker).	Workers AI, then HF DeepSeek, through our server with free daily quotas. Mock agents when AI is unavailable.
Docs	/docs	54 pages: user guides, architecture, operations, reference. Each one is also served as raw Markdown for AI agents.	—
NVXDEV	npx nvxdev init, pip install nvxdev	The DevStack CLI. It is reserved in the terminal signature but not published yet (hidden until the packages exist on npm and PyPI).	—

Commands are text only: our server never executes user, AI or template commands. Studio code runs only inside your browser (sandboxed preview iframe and PGlite).

เซิร์ฟเวอร์ของเรา ไม่รันคำสั่งใด ๆ ของผู้ใช้ AI หรือเทมเพลต โค้ดใน Studio รันในเบราว์เซอร์ของคุณเท่านั้น

Quick start / เริ่มต้น

Requirements: Node.js 22+ and npm. Studio's end-to-end tests also need Chromium (npx playwright install chromium).

git clone https://github.com/agents-sdk-space/devstack.bid.git
cd devstack.bid
npm install                       # also patches the OpenNext adapter (postinstall)
cp .dev.vars.example .dev.vars    # optional: CHAT_COOKIE_SECRET, HF_TOKEN, OPENAI_API_KEY (server-side only)
npm run dev                       # Next.js on http://localhost:3000

Production-style local run (Cloudflare Workers runtime, the same headers as devstack.bid). This builds the Studio, the docs bundle and the Worker:

npm run preview                   # http://localhost:8787  ->  open /studio/ and type "build a todo app"

Without any keys everything still works. The planner uses its offline rules, the Studio uses deterministic mock agents and placeholder images, and the chats report that they are unavailable.

ไม่มี key ก็ใช้ได้: planner ใช้กฎแบบออฟไลน์ Studio ใช้ agent จำลองและภาพ placeholder ส่วนแชตจะแจ้งว่ายังไม่พร้อมใช้งาน

Scripts

Script	What it does
npm run dev	Next.js dev server (:3000) with Cloudflare bindings
npm run studio:dev	Studio Vite dev server (:43123) with hot reload. It proxies /api to npm run dev.
npm run build	Docs bundle, then the Studio (public/studio/), then next build
npm run preview	OpenNext build plus wrangler dev (:8787): the real Workers runtime, headers and 404s
npm run lint / npm run typecheck / npm test	ESLint, tsc --noEmit, Vitest (site, API, quotas, Studio proxy and headers)
npm run studio:check	Studio brand sync, ESLint, tsc -b and Vitest (agents, PGlite + pgvector, export, static preview bundler)
npm run studio:e2e	Playwright against a running preview: the full "build a todo app" flow plus site regressions
npm run docs:check	Docs freshness: links, paths, env vars, scripts, manifest
npm run licenses / licenses:check	Regenerate / verify THIRD_PARTY_LICENSES.md (permissive licenses only; also enforced by npm test)
npm run export -- <id> --out dir	Writes setup.sh / setup.ps1 / README.md for a template
npm run deploy	Build plus deploy to Cloudflare. Needs owner approval (deploy guide).

The full list is in docs/reference/npm-scripts.md.

Architecture

Browser (no secrets)                                        Cloudflare Worker "nvx-stack-builder" (Next.js 16 via OpenNext)
───────────────────────────────────────────────────         ────────────────────────────────────────────────────────────────
/, /learn, /templates, /builder, /docs  (prerendered pages) ◄──  static assets + App Router, site CSP (no COEP)
/builder AI panel ──────────────────────────────────────►  POST /api/plan ─────────► HF DeepSeek V3 → offline planner
/chat ──────────────────────────────────────────────────►  /api/chat/turn ─────────► Workers AI (AI binding) ⇄ HF router
/chat/dev ──────────────────────────────────────────────►  /api/chat/dev/turn ─────► OpenAI Agents API (server key, low caps)
/factory (your key, per request) ───────────────────────►  /api/factory/* ─────────► OpenAI Agents API (your key)
/studio/  (static Vite app, COOP + COEP, Studio CSP)
  AgentOrchestrator → Code / DB / Design agents ────────►  /api/studio/llm/chat ───► Workers AI (Qwen Coder) → HF DeepSeek
  Static preview (Sucrase, sandboxed iframe) · PGlite ───►  /api/studio/image ──────► Workers AI FLUX
  Export (JSZip)                                            └ cookie identity (HMAC) · D1 nvx-db quotas · rate-limit bindings

• Stack: Next.js 16.4 App Router, React 19, TypeScript and Tailwind v4, deployed to Cloudflare Workers with @opennextjs/cloudflare. Storage is D1 nvx-db (chat history and quota counters). AI is the Workers AI binding plus the Hugging Face router.
• Studio: a separate Vite + React app in studio/ with its own lockfile. It is built into public/studio/ and served by the same Worker from static assets. COOP/COEP and the Studio CSP are scoped to /studio/* only; the preview page /studio/preview/ has its own no-network CSP. See docs/architecture/studio.md.
• Design system: a black, terminal-style "Apple developer" look with the NVX tokens (src/app/tokens.css), the animated pink-blob mascot, and the DevStack terminal signature window. The Studio receives a synced copy (npm run studio:brand). See docs/concepts/design-system.md.

Key handling / การจัดการ key

• The browser only ever calls our own /api/*. Server keys (OPENAI_API_KEY, HF_TOKEN, CHAT_COOKIE_SECRET) are Worker secrets set with wrangler secret put (stdin). They are never NEXT_PUBLIC_, never bundled and never logged, and error messages are redacted.
• The Studio has no key of its own. DEVSTACK's original LLM_API_KEY proxy was replaced by /api/studio/*, which uses the Workers AI binding and the existing HF token.
• Sandbox Factory and the planner accept your OpenAI key in the x-openai-key header of a single request. It is never stored.
• Every AI endpoint has a per-IP burst limiter (Workers Rate Limiting) plus daily quotas in D1 per signed cookie, per IP and site-wide. Quotas reset at midnight Bangkok time, and failed calls are refunded.

key ของเซิร์ฟเวอร์เป็น secret ของ Worker เท่านั้น ไม่เคยถึงเบราว์เซอร์ Studio ไม่มี key ของตัวเอง ทุก endpoint ที่ใช้ AI มีเพดานต่อนาทีและโควตาต่อวัน

Fallbacks / ทางสำรอง

When	What happens
No OPENAI_API_KEY / HF down	The planner uses its offline rule-based plans
A Workers AI model fails	Chat and Studio move to HF DeepSeek V3 (and HF to Workers AI)
No AI or D1 on the server, or the Studio quota is used up	Studio agents use deterministic mock LLM, template schema/app and placeholder images. The chat shows CodeAgent (mock) etc.
Generated code does not compile or imports a package other than react / react-dom	The preview shows the error and the Output tab names the file. Agents, DB, assets and export keep working.
R2 not enabled	Chat attachments return uploads_disabled

Deploy

devstack.bid runs on Cloudflare Workers (Worker nvx-stack-builder, custom domains devstack.bid + www). Deploying, pushing, merging to main, DNS changes and new Cloudflare resources all require the owner's approval (see AGENTS.md).

npm run lint && npm run typecheck && npm test && npm run docs:check && npm run studio:check
npm run cf:build                 # includes the Studio build (public/studio, ~34 MB, largest file 10 MB < 25 MiB limit)
npm run deploy                   # approval required

Step-by-step guide, secrets, D1 migrations and rollback: docs/operations/deploy-cloudflare.md · docs/operations/env-secrets.md · docs/operations/rollback.md.

Contributing / ร่วมพัฒนา

We welcome issues and pull requests. Please read CONTRIBUTING.md and the Code of Conduct. To add a template, follow docs/contributing/add-template.md. Keep both languages in every UI string and update the docs in the same change.

เรายินดีรับ issue และ pull request ทุกขนาด อ่าน CONTRIBUTING.md ก่อนเริ่ม ทุกข้อความใน UI ต้องมีทั้งไทยและอังกฤษ และอัปเดตเอกสารใน PR เดียวกัน

License

MIT © NVX DEVELOPER BOTH and contributors. The Studio is the team's DEVSTACK workspace (v0.2.0), merged into this repository under the same license. Every dependency is open source under a permissive license (MIT, ISC, BSD, Apache-2.0, OFL-1.1 for fonts; PGlite/pgvector under Apache-2.0 / the PostgreSQL License) with no commercial or usage-based terms: see THIRD_PARTY_LICENSES.md. Tool logos come from Simple Icons (CC0) and are trademarks of their owners (NOTICE).