A personal productivity system powered by Claude Code. Clone this repo, open it with claude, and the AI guides you through setup.
Compass gives you an AI-powered chief of staff, personal coach, and communications drafter — all backed by a local knowledge base that learns about your work over time.
It combines:
- kbx — a local-first knowledge base that indexes your meetings, notes, people, and projects with hybrid search (full-text + semantic)
- cc-marketplace plugins for Claude Code:
- chief-of-staff — meeting prep and debrief, daily briefings, task management, decision tracking, strategic advisory
- inner-game — personal coaching, journaling, life assessment (based on Tim Gallwey's Inner Game)
- draft — voice-aware message drafting for Slack, email, and announcements
- Claude Code (CLI)
- Python 3.11+
- uv (Python package manager)
git clone https://github.com/YOUR_USERNAME/compass-template.git my-compass
cd my-compass
bash setup.shThe setup script installs kbx (with semantic search and embeddings), builds the search index, registers the plugin marketplace, and installs the Claude Code plugins.
Then open Claude Code:
claudeClaude will detect it's a fresh install and guide you through:
- Setting your name in
kbx.toml - Running
/cos:setupto connect your task manager, calendar, and chat tools - Optionally setting up the coach (
/inner-game:setup) and drafter (/draft:setup)
| Tier | What you get | Requirements |
|---|---|---|
| Tier 0 | Knowledge base + tasks.md | kbx only (works immediately) |
| Tier 1 | + task management via your preferred tool | Connect a task CLI or MCP |
| Tier 2 | + Slack, email, calendar, project tracking | Connect MCP servers via claude.ai |
You start at Tier 0 and add capabilities as you connect more tools. Everything degrades gracefully — you never need all of them.
All your data stays in this repo under memory/:
- Meeting transcripts and notes
- People and project profiles
- Decision logs and tagged notes
- Journal entries and coaching files
- Company context and glossary
The search index (kbx-data/) is gitignored — it's rebuilt locally from your files.
This repo is a blank slate. All personalisation happens when you run the setup commands. Your data lives in memory/ and never leaves your machine unless you push it.
Your AI operations partner. Handles meeting lifecycle (prep → attend → debrief), daily briefings, task management, decision tracking, and strategic review.
Key commands: /cos:setup, /cos:briefing, /cos:prep, /cos:debrief, /cos:todos
Personal life coach based on Tim Gallwey's Inner Game framework. Coaching conversations, daily journaling, identity work (The Document), and life domain assessment.
Key commands: /inner-game:setup, /inner-game:morning, /inner-game:evening, /inner-game:session, /inner-game:document
Communications drafter that learns your voice. Creates messages for Slack, email, and announcements in your authentic style, adapted to the audience.
Key commands: /draft:setup, /draft:draft, /draft:restyle
Compass includes a built-in multi-agent team for the Chief of Staff plugin. Three persistent agents run in separate tmux panes, each specialising in a different domain.
| Agent | Role | Owns |
|---|---|---|
| ops | Daily operations, tasks, accountability | /cos:briefing, /cos:todos, /cos:status, /cos:decision |
| briefer | Meeting lifecycle — prep and debrief | /cos:prep, /cos:debrief |
| advisor | Strategic advisory, coaching, patterns | /cos:review, /cos:coach, /cos:blindspots, /cos:culture, /cos:codify, /cos:supergoal |
/cos-boot — spawn all 3 agents in a tmux layout
/cos-focus <agent> — spotlight ops, briefer, or advisor (resize panes)
/cos-focus reset — restore default layout
/cos-shutdown — graceful shutdown, persist state to kbx
The agents coordinate automatically — ops delegates meeting work to briefer, flags patterns to advisor, and synthesises their input. Each agent reads and writes to the shared kbx knowledge base so context survives across sessions.
Layout (default):
┌───────────────┬──────────┐
│ │ briefer │
│ ops (~45%) ├──────────┤
│ │ advisor │
└───────────────┴──────────┘
- Writing CLAUDE.md files — how to structure CLAUDE.md for maximum agent effectiveness (synthesised from HumanLayer, Vercel's eval findings, and production patterns)
This is a template — fork it, make it yours. If you find improvements that would help everyone, PRs are welcome.
MIT