Pi + Ollama: Qwen 3.6 coding agent setup

This tutorial pairs the Pi coding agent with Ollama to build a restaurant website using Qwen 3.6. The presenter argues that a lean coding harness leaves more context available for the task than Claude Code. His context overhead figures and quality thresholds are claims made in the video, rather than universal benchmarks.

Setup covers installing Pi through npm, downloading a model with Ollama, and adding provider entries to .pi/agent/models.json. LM Studio appears as another local inference option. For Qwen 3.6, the presenter gives approximate VRAM requirements of 16 GB for the 27B model and 24 GB for the 35B model. These are his sizing estimates.

The workflow uses agents.md for project rules and design.md for styling. The presenter recommends ChatGPT or Claude for a detailed implementation plan, then splits that plan into ordered feature files. Pi implements each feature in a fresh session to limit context growth. This makes the coding assistant part of a mixed workflow that also uses hosted services.

The demonstration adds Pi Web Access for search and image retrieval; the presenter describes MCP support as an extension rather than a default capability. The resulting site includes a responsive menu and category filters. Deployment uses a zipped public_html folder uploaded to Hostinger. The tutorial therefore covers local AI coding alongside web access and hosted deployment, rather than an entirely offline process.