~ / projects / chief-of-staff
11/14

Kieran's Chief of Staff

LIVE SINCE 2025-12 LATEST 2026-05-15 CATEGORY · PERSONAL INFRA
WHAT IT DOES§ 01

A phone-first chief of staff — runs the day across diary, comms, knowledge recall, and on-call infra fixes, calibrated to my voice and my views. Cherry-picks the best ideas from Hermes Agent and OpenClaw (persistent layered memory, autonomous skills, messaging-first interface, scheduled tasks, subagent delegation) and builds them directly on Claude, so it integrates from day one with my wiki, prompt library, and 18 MCP servers — no generic agent framework in the middle.

Voice-message it on the school run and get a transcribed answer back; photo a whiteboard and ask for a summary; queue a 10-min background agent and carry on. Pre-meeting alerts, morning briefings, mid-day check-ins, and evening wrap-ups happen without me asking. The longer it runs, the more it represents how I'd actually answer.

HOW IT'S BUILT§ 02

Bridges to Claude Code via the Claude Agent SDK, so anything I can ask Claude from a terminal I can ask from anywhere. Text, voice (local Whisper), photos (local GLM-OCR), and documents in; streamed responses with tool-status emoji and file auto-send out. Smart model routing across Haiku, Sonnet, and Opus with per-model cost tracking, and skill-triggered model bumps for harder tasks.

Local-first inference where possible: Whisper, Qwen 3.6 35B-A3B tool-calling, Qwen3-Embedding, and GLM-OCR all run on the Local LLM Stack, keeping morning briefings and routine queries at zero Anthropic cost. Proactive layer schedules reminders, briefings, check-ins, wrap-ups, and an Obsidian vault task watcher.

Calendar, Reminders, and Email moved off AppleScript to Home Assistant REST in April 2026 (TCC silently lost access too often) — event-driven email reactions are filtered by a local Qwen smart filter and routed via HA → bot webhooks. Autonomous background agents queue one-at-a-time with a 10-min cap and a multi-subtask orchestrator. Knowledge base with FTS5 + local vector search auto-indexes conversations, agent results, and workflow outputs. Connects to 18 MCP servers. macOS LaunchAgent with 10-layer security and a subprocess priority queue.

TECH STACK§ 03
  • Claude Agent SDK
  • Bun
  • TypeScript
  • Grammy
  • MCP (18 servers)
  • Local Whisper
  • Qwen 3.6 35B-A3B
  • GLM-OCR
  • SQLite + FTS5
  • Home Assistant REST
  • Telegram API
  • macOS LaunchAgent
METRICS§ 04
18MCP SERVERS
£0ANTHROPIC · LOCAL QUERIES
10SECURITY LAYERS
1SUBPROCESS QUEUE