Source: https://youtu.be/rVzGu5OYYS0 — "ClaudeClaw" personal AI OS walkthrough Date: 2026-04-17 Context: Ticket MC-655 — identify features worth adopting into Luci/PKA/Mission Control
ClaudeClaw is a Mac Mini-hosted personal AI OS built on Claude Code + Agent SDK with 5 named agents, a voice "War Room", unified memory, and multi-channel bridges. We already have ~70% of the architecture. The genuinely new ideas worth stealing:
| ClaudeClaw feature | Luci equivalent |
|---|---|
| Mission Control dashboard | luci-dashboard.service on port 3001 |
| Telegram/WhatsApp/Slack bridges | Telegram MCP + WhatsApp Monitor |
| Claude Code + Agent SDK core | Same (Luci runs Claude Code persistent) |
| launchd services | systemd services |
| Cloudflare Tunnel for remote access | Tailscale + direct IP |
| Scheduler / cron missions | scheduler.py |
| Obsidian vault injection | PKA/Vault + auto-memory |
| SQLite memory | vault.db |
| Worker ticket queue | MC ticket system |
| Audit Log | vault.db activity_log |
| Chat ID Allowlist | Telegram allowlist |
1. War Room (voice stand-ups) — ★★★★★ Live voice boardroom where Elmar talks to agents Google-Meet-style. Stack: Pipecat (Python orchestrator) + Gemini Live (speech-to-speech) + Silero VAD (voice activity detection). Fallback legacy stack: Deepgram STT + Cartesia TTS. Why it matters: Turns "type Telegram messages" into "hop on a call with Luci while driving". Massive UX upgrade over text. Effort: Medium. Pipecat is open source, Gemini Live API we already use.
2. Named specialist agents — ★★★★ Instead of one Luci handling everything, split into Main (triage), Comms, Content, Ops, Research. Each has its own chat, own memory slice, own personality. Why it matters: Current Luci is a generalist; context-switching between deep-research and email-drafting pollutes context. Named lanes = cleaner prompts. Effort: Low-Medium. Mostly naming + routing convention. systemd template units can launch N workers.
3. Smart classifier router — ★★★★ AI-decides which agent handles a message instead of relying on prefixes/keywords. Rules demo'd: - "everyone/team/status" → broadcast - "@comms ..." → route to comms - Pin file exists → route to pinned agent - Default → Main Why it matters: Friction reduction — no more remembering "is this a /research or /work command". Effort: Low. One Haiku call per inbound message.
4. Hive Mind unified memory — ★★★ All agents share a completed-tasks ledger + preference/fact store. Gemini 768-dim embeddings for semantic, FTS for lexical, SQLite for persistence. Why it matters: Right now each Luci session has its own memory; if Comms learns something, Ops doesn't see it. Comparison: We already surveyed this landscape (2026-04-11 memory reports). A Gemini-embedding + FTS hybrid is exactly what those reports recommended. Effort: Medium. We have the research; needs a build session.
5. Meeting Bot with pre-flight briefing — ★★★ Agent joins Daily.co/Google Meet calls, having pre-read Gmail thread + calendar context for the meeting. Optional Pika video avatar. Why it matters: High-leverage for Elmar's actual work (lots of Safair/partner calls). Currently he has to brief each call himself. Effort: Medium-High. Daily.co + Gmail API + calendar already available, but avatar and TTS loop is new.
6. Frame-based pipeline pattern (Pipecat) — ★★★ Every event is a "Frame" (audio chunk, transcript, LLM response) flowing through a left-to-right pipeline of processors (VAD → STT → LLM → TTS → Transport). Composable and inspectable. Why it matters: Pattern worth copying even outside voice — useful for any streaming workflow (Telegram message → classifier → agent → memory-write → reply). Effort: N/A — this is a design pattern, not a library to install.
7. Kill Phrase + PIN Lock — ★★ Nice-to-have security. We have Chat ID allowlist which covers 95% of the threat model. PIN adds friction for little gain on a personal device.
8. Cost Footer in replies — ★★ Show token spend per reply. Cute but noisy. We can track via logs if we care.
9. Pika video avatar — ★ Avatar in Meet calls. Cool but gimmicky; voice is enough.