Measure the GPU offload instead of estimating it, and keep the development scratch out of the repository

Offload what the card can hold, survive a dropped connection, and stop sending prompts the model cannot read

Workers print their own code and are claimed from the site; fetch and verify llama.cpp

Harden every component declaration against Gradio's reset; denser library grid

Drawn mockups instead of recorded demos; delete conversations; drop Llama; fix the weights cache re-downloading

Say how a worker is shared: one access code, any number of people

The pairing code appears when you open the panel; one size for section headings

Real setup steps instead of a download that never existed; mobile notice, veil and control fixes

Library as its own page; cards show what each member makes; Anthropic-calm palette; artifact previews in answers

Library page with tested demos; no-agent veil; energy meters and J/token; 15 new skills; external MCP catalogue; py3.10 gate