parrot: a minimal macOS dictation daemon. Hold Fn, speak, release - text appears at your cursor. Whisper on the Apple Neural Engine, fully on-device, running as a login daemon - plus a hotword hack that routes āJarvis, ā¦ā straight into Claude Code.
My siteās chat agent answers with exact deep links to my posts - no search tool, no vector database. It reads /llms.txt at prompt time, and the whole āretrieval pipelineā is one HTTP GET.
Wire an xpander agentās success/error webhook into hark.ryan.ceo and your phone buzzes the moment a long-running agent task finishes - no polling, no forgotten tabs.
The āDavidās AIā sidebar is a real xpander agent - created with one API call, streamed with another, and dropped onto a static site behind a tiny key-holding Worker.
A static Hugo site on GitHub Pages with zero backend - except one small Cloudflare Worker for the chat, an llms.txt for the agent, and a calendar in the sidebar.
Home Assistant OS on a Raspberry Pi 4, seven Lutron CasƩta blinds driven by the Adaptive Cover integration, and an OpenClaw agent living on the Pi as an add-on.
How I run my own Google Photos: Immich on a Graviton EC2 box, originals on an S3 rclone mount, hot data on EBS, private over Tailscale, with a real Letās Encrypt cert and zero open ports.
The more intelligent the runtime becomes, the more expensive it will be to leave.
They donāt give you an AI stack. They give you dependency.
Instead of full delegation, we are literally working harder, not smarter.