Qwen 3.6 + Pi: Set up a local Zendesk support agent

This tutorial builds a customer support AI agent with Pi, an open source harness, and Qwen 3.6 35B running through LM Studio on the speaker's Mac Studio. The Pi LM Studio extension connects the harness to the local LLM.

The speaker asks Pi to read its own documentation and create a shareable agent package with a Zendesk extension, response prompts and skills. Zendesk credentials go into a .env file. The speaker reports successfully listing 33 tickets and sending a reply to ticket 115 after Pi fixes how the extension reads those credentials. During setup, the model also gets stuck in a loop that requires cancellation.

Automation uses Pi's print mode, -p. The speaker requests a cron job and polling script that checks for tickets with status new every five seconds, generates a response and sends it through the Zendesk API. A test ticket receives a troubleshooting reply. Model inference runs locally, but the workflow still communicates with Zendesk.

Business-specific FAQs and skills still need work. The speaker explicitly leaves evaluations, personal information redaction, prompt injection defenses and cloud deployment outside the tutorial. Gmail and Outlook are suggested as other inbox targets; Zendesk is the demonstrated integration.