Agent Global Hub

A personal infrastructure layer for Claude Code — centralizing plugins, skills, hooks, and project bootstrapping across a multi-project workspace.

Why

Claude Code is powerful but each project is an island. This hub provides the shared code layer: a brain MCP for cross-project memory, plugins that inject context at session start, a registry that tracks all projects, and bootstrapping for new ones. (Global config + skills live natively in ~/.claude/.)

Features

• Plugins: haint-core (session hooks, context injection), godot-dev (Godot workflows)
• Brain-powered memory (haingt-brain MCP) for cross-project context
• Project registry with drift detection
• Bootstrap script for new projects

Global config + skills (CLAUDE.md, settings.json, skills/, brains/) live natively in ~/.claude/ — not in this repo.

---

Note: All agent rules and configs are per-project. This hub only contains shared scripts, templates, and plugins.

Structure

~/Projects/agent/
├── bin/
│   ├── bootstrap-project.sh   # Bootstrap new project + auto-register in hub
│   ├── ag-registry-audit.sh   # Full drift check: registry vs actual state
│   └── shell-aliases.sh       # Shell shortcuts (source in ~/.zshrc)
├── registry.json              # Orientation catalog of all child projects
├── .claude/scripts/
│   └── registry-check.sh      # SessionStart hook: lightweight drift alert
├── plugins/
│   ├── haint-core/            # Core plugin: hooks (SessionStart, PreToolUse, PreCompact, ...)
│   └── godot-dev/             # Godot plugin: gdformat, GDScript workflows
├── mcp/
│   └── haingt-brain/          # Custom MCP server: semantic memory + knowledge graph
└── templates/
    └── .claudeignore          # Default ignore template for new projects

Per-Project Structure

Every project has this structure (created by bootstrap):

project/
├── .claude/
│   ├── CLAUDE.md           # Project context, values, conventions, security
│   ├── settings.json       # Project-specific hooks (if any)
│   └── skills/<name>/      # Skills (SKILL.md + supporting files)
└── .mcp.json               # Project-level MCP servers (where needed)

Architecture

Token Optimization

Rules are minimized to reduce per-turn token cost:

What	Where	Token cost
Enforcement (security, dangerous commands)	settings.json hooks	0 (runs as shell)
Core directives (no dirty state, reversibility)	.claude/CLAUDE.md Values	Once per session
Brain context (decisions, prefs, last session)	SessionStart hook output	Once per session
Project-specific workflows	skills/<name>/SKILL.md	On invocation only

Skills (.claude/skills/<name>/SKILL.md)

Skills use YAML frontmatter for invocation control:

• All skills are invocable by Claude when relevant or by user with /name
• Supporting files (templates, references) live alongside SKILL.md in the same directory

Hooks (settings.json)

Hook	Purpose
SessionStart	Inject git status + brain context
PreToolUse (Bash)	Block dangerous commands, scan for secrets in staged files
Stop (project-specific)	Auto-format on save (e.g., gdformat for Godot)

Project Registry

registry.json (v2) is an AI/human orientation catalog of all child projects — {path, type, summary, status?} per project: the facts the filesystem can't cheaply reveal. Capability lists (skills/plugins/rules/mcps) are derived on demand (ls .claude/skills, jq .mcp.json, jq installed_plugins.json) and deliberately not stored — index_tools.py already derives them from the filesystem, so a stored copy is redundant token-rent that drifts. status is omitted when active (default); set to primary/postponed/archived otherwise.

Component	Role
registry.json	Orientation catalog: path, type, one-line summary, lifecycle status
bin/ag-registry-audit.sh	Audit: every dir registered? every registered path exists?
.claude/scripts/registry-check.sh	SessionStart hook: silent when clean, flags NEW/STALE dirs
bin/bootstrap-project.sh	Registers a slim stub on bootstrap; /project-creator fills the summary

Drift detection (audit + SessionStart hook share bin/registry-lib.sh so the guard can't itself drift) flags three things:

• NEW — an unregistered ~/Projects/<dir>.
• STALE — a registered path that's gone.
• NARRATIVE-SYNC — the project the registry marks status:primary must be the one the always-loaded narrative surfaces crown: indie-ecosystem.md ("center of gravity" row) and core-memory.md ("PRIMARY BUILD = …" line). A pivot that updates the registry but forgets the narrative (or leaves two primaries) is flagged deterministically — so the canonical status and the loaded context can't silently desync.

A human resolves flags; registration is via bootstrap-project.sh + /project-creator. No auto-sync/self-healing.

Quick Commands

ag-help          # Show all commands
bootstrap <dir>  # Setup new project
ag-status        # Check agent setup across all projects
cdc <project>    # Switch to project directory

MIT License