Three internal tools the portfolio runs on.

The planning database every other tool reads from.

Vite admin dashboard on top of a Firestore that holds epics, items, and milestones across 18 workspaces. The same database is exposed as an MCP server, so Claude Code, agents, and the Product Partner dogfood instance read the live roadmap directly.

• Hierarchical planning model: workspace → project → epic → item
• MCP server with read and selective write tools for agent clients
• Same Firestore powers both the admin UI and the agent-facing adapter
• Activity feed + cleanup scripts so the planning layer doesn't drift

An Obsidian vault that synthesizes raw inputs into nine queryable wikis.

Raw sources (Claude conversations, PDFs, LinkedIn exports, source repos) land in `raw/` and get synthesized into structured `_index.md` pages. ClaudeClaw agents write scheduled outputs straight into the right `raw/` directory. Downstream tools read the synthesized layer.

• Nine wikis: writing, BookOwl, self, tech, reading, career, networking, work-history, consulting
• Raw inputs and synthesized output sit in different folders, not mixed together
• CLAUDE.md schema so any model working in the vault knows where things go
• Editable from a phone via iCloud sync

A multi-agent system that runs the boring, recurring ops work.

Claude Code plugin with a persistent daemon, cron-style scheduling, and a trust engine in front of every action. A small Haiku watchdog polls and decides what's worth escalating to a flagship model. Notifications route by urgency, not by source.

• Orchestrator auto-discovers agent groups from `config/groups/`
• Per-agent trust levels: autonomous, notify, or approval-required
• Watchdog on Haiku keeps polling cheap; flagship models only run when invited
• Scheduled outputs flow back into Workspaces for later synthesis