An AI customer-service agent that knows when to tap out

What it does

Customer-Agent is a PyQt6 desktop app for e-commerce sellers, currently wired to Pinduoduo via WebSocket. An LLM agent answers buyer chats autonomously: it can pull the shop’s product list, send product cards, look up product details, and search a customer-service knowledge base covering returns, shipping, and refunds. When a buyer’s intent trips configured keywords, it transfers the conversation to a human.

The interesting bit

The agent framework is self-built rather than borrowed — multi-turn tool calls plus session context, with two separate knowledge bases (product vs. policy) so product facts and refund rules don’t blur together. Product knowledge is even auto-ingested by sending Pinduoduo listing text and images to a multimodal LLM.

Key highlights

• Five agent tools, from get_shop_products to transfer_conversation
• Async message queue with a handler chain, plus WebSocket auto-reconnect and heartbeat
• LLM access via LiteLLM with six preconfigured Chinese providers and an OpenAI-compatible option
• Endpoint hygiene: HTTPS enforced, local endpoints require explicit opt-in, keys stored with Windows DPAPI
• Ships as a ~100 MB installer with bundled Playwright driver; no admin rights needed

Caveats

• One platform only (Pinduoduo) and Windows-only; the README doesn’t hint at other channels yet
• Product sync sends listing text and images to your chosen LLM vendor — data retention is your problem to verify, and the README says so plainly
• Config changes only apply to accounts restarted after saving, which is the kind of quirk you discover the hard way

Verdict

Worth a look if you run a Pinduoduo shop and want an auto-responder with a sane escape hatch to humans. If you’re hoping for a multi-channel, cross-platform bot, this is a single-storefront tool with clear ambitions and clear limits.