Nagi
Nagi is an AI elder-care product where the AI's job is to route routine requests and to step aside the moment a human family member should take over — and where that family member can self-host the entire stack on a $5 s
Core trick
Routes routine elder requests to the model and hands escalations to a family member on a self-hosted stack.
Description
Nagi is an AI elder-care product where the AI's job is to route routine requests and to step aside the moment a human family member should take over — and where that family member can self-host the entire stack on a $5 server they trust or own.
The world is aging, and there are fewer younger family members available to provide care. It's a problem that grows every year, but one we can start preparing for now.
Nagi is built on three capabilities that only recently became viable at consumer level: 1M-token context, prompt caching, and LLM-quality voice conversation — all wrapped in privacy and elder-paced interaction, so trust isn't an afterthought.
We're building for what's next because the technology already crossed the threshold that makes this work, and most builders haven't noticed yet. The same architecture that lets one family run Nagi on a $5 server lets governments and NGOs digitize their pro-bono outreach at low cost . The goal is reducing isolation for the elders families struggle to reach; and for those with no family to reach them at all.
README excerpt
# Cedar / Nagi (凪) > By 2030, over **1.4 billion people** will be 65 or older — three out of every ten adults in many countries. Most of the apps their families take for granted weren't built for them. The result is a quiet exclusion: from grandchildren's photos, from medical portals, from the ordinary digital fabric of a family's life. **Nagi** is open-source infrastructure for closing that gap. Not a locked-down launcher, not a panic-button device — an AI-native companion the elder talks to, configured by the people who love them. **Cedar** is the project. **Nagi** (Japanese 凪 — the calm after the storm) is the product. --- ## For Anthropic 4.7 hackathon judges The shortest path to feeling what Nagi does: - **What's 4.7-native here:** `claude-opus-4-7` powers the elder conversation. With `profile.long_context_recall = true`, the chat path loads up to ~200 prior turns from `activity_log` into the system block — Opus 4.7's 1M context lets Nagi recall what the elder said weeks ago without RAG, embeddings, or a vector store. See [`supabase/functions/ai-chat/index.ts`](supabase/functions/ai-chat/index.ts) — the `Long-context recall` block. The same model powers the family-facing weekly digest in [`supabase/functions/generate-digest/index.ts`](supabase/functions/generate-digest/index.ts). - **Two more 4.7-line capabilities in the same file:** three-block prompt caching (static policy → skills → per-elder profile) preserves the warm-path cache when the long-context block is