AgentHandover turns screen habits into Skills

This is one of the more credible “teach by watching” attempts because it treats observed behavior as a source of reusable procedure, not just a transcript dump. The real test is whether it can generalize intent cleanly enough to avoid locking in accidental clicks or noisy habits.

• –The local-first stack is the strongest part: on-device vision, local embeddings, and encrypted-at-rest storage make the privacy story believable.
• –MCP integration is the right distribution layer; if the Skills are genuinely structured, they can travel across Claude Code, Cursor, Codex, and other agents without bespoke glue.
• –The self-improvement loop is the differentiator: confidence scores, deviation tracking, and failure-based demotion are more useful than static SOP export.
• –Passive discovery is also the riskiest piece, because repetition is not the same as intent; without strong review UX, it could learn brittle or misleading patterns.
• –Mac-only and model-quality constraints mean this lives or dies on Apple Silicon performance and how well Gemma 4 handles screen understanding in the wild.