An e-commerce operations agent, run from Telegram
The problem
Running a small e-commerce brand means a lot of small, repetitive, context-heavy tasks spread across different tools: tracking who ordered what and when for B2B reorders, checking recent Shopify orders and sales, marking orders fulfilled after posting them, and keeping production and inventory data current — all manual, all context-switching.
The goal: put one conversational interface in front of all of it, so these tasks happen by asking.
What it does today
- Shopify (read + write) — pulls recent orders by country, gives a sales snapshot, and marks orders fulfilled directly from a Telegram message.
- Daily morning brief — a scheduled summary of the last day’s / week’s Shopify situation, sent automatically via cron.
- Google Drive (read + write) — reads and updates company files, including parsing vendor order-confirmation forms for B2B reorder tracking.
- Notion (read + write) — maintains a company knowledge base built by ingesting Drive into Notion.
- Trello (write) — updates and manages tasks.
- Production sheet — reads and appends new rows on instruction, replacing a manual update step.
How it’s built
Built on OpenClaw — an open-source, self-hosted, model-agnostic agent framework — running on a repurposed laptop as an always-on Linux server.
| Layer | Choice |
|---|---|
| Interface | Telegram (text + voice) |
| Reasoning | Model-agnostic — currently Google Gemini Flash, after testing OpenAI Codex and Claude (cost-driven decision) |
| Short-term memory | OpenClaw’s built-in memory — .md instruction files defining what to persist and how to reason per task type |
| Long-term memory | A Notion knowledge base, built by ingesting Google Drive — a curated retrieval layer, deliberately not a formal vector-RAG pipeline |
| Custom build | A Shopify app built from scratch to give the agent authenticated store access |
| Skills | Community skills from openskills.org, downloaded and tuned |
The honest part
WhatsApp access never worked; Telegram eventually did. But the real wall was agent quality — memory, skills, and reasoning. Out of the box the agent felt “stupid”: it needed continuously more context and explicit rules on how to think about each task type. Adding those rules also pushed it toward overengineering simple tasks — a tuning trade-off still being worked out.
Getting an agent connected is the easy 20%. Getting it reliable and appropriately-scoped is the hard 80%.
What’s next
- Move hosting off the laptop to a low-cost, always-on cloud host
- Add guardrails and permission scoping before rolling out to other users
- Extend access to the co-founder and employees
- Add email integration, a working WhatsApp channel, and an inventory dashboard