Hibiscus

Inspiration We wanted to build something that could actually do things — not just chat. Most AI tools stop at generating text. We wanted an agent that could pick up a plain-English task and carry it all the way through: writing a post, generating an image, sending a Telegram message, updating a Shopify listing. We called it Hibiscus.

What it does Hibiscus is an autonomous multi-step operator agent. You give it a prompt; it plans a sequence of tool calls and executes them end-to-end without hand-holding. It can generate images (Qwen via Chutes), produce videos (Replicate), draft and publish Substack posts, send Telegram messages, update Shopify products, and upload TikTok videos — all from a single input. The frontend shows you a live workflow stepper so you can watch each phase execute in real time.

How we built it The agent logic runs entirely on Convex — a reactive backend that lets us invoke long-running actions without managing servers or queues. The orchestration LLM (Azure GPT-5.4) plans tool calls and dispatches them through the AI SDK. Each external integration is its own Convex action. The frontend is a Vite + React SPA with Convex Auth for user and org management, and ShadCN for the UI.

Challenges we ran into Getting a single agent to reliably chain heterogeneous tool calls — image generation → Telegram delivery → Shopify update — without losing state between steps was harder than expected. Convex's action timeout model forced us to think carefully about which steps could be parallelized and which had to be sequential. We also had to wrangle several competing scaffolds and framework artifacts that crept into the repo during rapid iteration.

Accomplishments that we're proud of A real end-to-end pipeline: one text prompt triggers planning, multi-tool execution, and structured result delivery — all observable from the UI. We also built a clean multi-tenant org system on top of Convex Auth with role-based access, invite flows, and password reset — from scratch.

What we learned Convex is genuinely excellent for agentic workloads — reactive queries, durable actions, and built-in auth compose surprisingly well. We also learned that keeping generated code (VibeFlow produced 2,100 lines of actions) auditable matters early, not as an afterthought.

What's next for Hibiscus Persistent memory across runs, scheduled/cron-triggered workflows, and a visual flow editor so non-technical users can wire up their own tool chains without writing a prompt. We also want to add more output channels — email newsletters, LinkedIn, and webhook-driven triggers from external events.

Built With

• ai-sdk
• azure-openai
• browser-use-api
• chutes-(qwen)
• convex
• convex-auth
• openai-api
• react
• replicate
• resend
• shadcn-ui
• shopify-graphql-api
• tailwind-css
• telegram-bot-api
• typescript
• vite