From hackathon prototype to production SaaS — RED gives you the Agentic-Native stack in 90 minutes. So your next 3 months build the product, not the plumbing.
Server functions, ACID transactions, CQRS, realtime subscriptions. Every feature of the AI Module is a thin, type-safe layer on top.
Most "AI starter kits" give you a streaming chat input and call it a day. RED ships five production-grade primitives built on Convex's reactive database.
Everyone in a thread sees the same tokens, the same typing indicators, the same task updates — in real time. Built on Convex's reactive queries, not a bolted-on socket server.
Users type while the agent streams. Messages land in an ordered queue, processed with back-pressure. No lost turns, no race conditions, no rebuilds.
A Boundary is a first-class container: dedicated memory, scoped knowledge base, members, permissions.
Tools, memory scopes, models, guardrails, fallbacks — declared, not wired. Spin up a new agent in a file.
agent("researcher", {
model: "gpt-4.1",
tools: [search, scrape],
memory: "boundary",
tasks: [deep_research],
}) Quick tool calls that finish inside the turn. Standard chat affordances. Nothing exotic.
Sync tasks finish inside a turn. External tasks run for hours, off the main thread, and push progress back in real time. Your user closes the tab. Your agent keeps going.
Deep research. Batch scraping. Multi-step agent pipelines. Video renders. Fine-tunes. Things LLM wrappers can't do.
task("async", …) — runs forever, pings on done.bun install.Most starter kits try to please everyone — five databases, three ORMs, four CSS systems, ten abstraction layers. You spend the first week deleting what you don't need, and the second week realizing you can't.
RED picks a side. One stack, chosen on merit. Open the codebase and change anything — nothing is hiding.
One runtime. One database. One router. One component library. Picked because they're the best at their job today — not because we wanted to sell optionality.
RED isn't a SaaS kit with AI glued on. The AI Module is the spine — auth, billing, i18n, routing all wrap around it.
Buy the license, drop in the key, run the CLI. No GitHub invites, no repo to clone, no setup wiki.
$ bun create red-app my-saas \ --license YOUR_LICENSE_KEY › verifying license ………………………… ok › pulling RED@v1.4 › scaffolding ./my-saas › installing deps with bun › provisioning convex dev deployment › seeding boundaries + demo agents ✓ running on http://localhost:3000 $ cd my-saas && bun red update › v1.4 → v1.5 · 6 conflicts auto-resolved
Validates your license, pulls the latest tagged release, scaffolds the project, wires Convex, and starts the dev server.
Reads the diff between your release and the new one, applies it to your codebase, and resolves conflicts with your own edits.
red-init ·
red-migration built for the agent workflow
Ship to a generous cloud free tier today. Or self-host on your own box. Same codebase, either way.
Backend on Convex (reactive DB, server functions, cron, file storage). Frontend on Cloudflare Pages (global edge, free SSL, preview URLs).
Convex is open-source, so is everything else in the stack. Docker-compose onto a VPS, an air-gapped server, or bare metal.
$ docker compose up -d › convex backend :3210 › app (vite/bun) :3000 › postgres :5432 ✓ RED running on-prem
One-time license per use case. No seats. No usage tiers. Source included.
For your own products.
For teams and agencies.
100M+ tokens burned on GPT-5.4 XHIGH and Opus 4.6/4.7 MAX. Months of carefully designed modules and security hardening, backed by 20 years of coding.
You will not one-shot this quality. You'll spend weeks on an inferior result — which is exactly why I built RED. The baseline here is first-grade SaaS craft: the kind you see in Notion, Airtable, Linear.
Skip the burden. Push the RED button.
If your product runs on an LLM today — or might in the next 18 months — yes. Every SaaS is becoming agentic, it's a matter of when, not if. RED isn't a SaaS kit with AI bolted on. It's a kit where the AI Module is the spine: auth, billing, i18n, routing all wrap around it. You get the primitives (streaming, queue, boundaries, async tasks) exactly when you need them.
Every hour wiring websockets is an hour your competitor spent talking to customers.