Pi setup with Kimi K2.6 and custom Archon extensions

This live setup explores Pi as a minimal coding assistant that users adapt through extensions and configuration. The speaker contrasts that approach with Claude Code and Codex, arguing that Pi needs more setup but gives him more control over his workflow.

The session connects Kimi K2.6 through Kimi for Coding using an API key, installs Pi Web Access from npm, and tests an extension package with a status line, permission gates and a questionnaire. During verification, the agent prints the API key in the terminal; the speaker deletes it and creates a replacement. He also cautions that the extension catalog is not officially vetted.

Skill loading takes some troubleshooting. Copying skills into .agents/skills works in the demonstration. The speaker then configures .pi/settings.json to load the existing .claude/skills directory and confirms discovery after a reload.

Most of the build concerns a custom Archon dispatch extension for background workflows. It adds approval prompts and a running-workflow indicator, but testing exposes branch handling and output capture problems on Windows. Workflows eventually launch successfully, while returning their actual results to the Pi conversation remains unresolved at the end.

The local LLM discussion is exploratory: the speaker considers quantized Qwen 3.6 models for his RTX 3090 hardware but does not test them here. The demonstrated setup uses a remote model service.