1. claude-agent-catalog — agent inventory (start the story on GitHub)
2. claude-marketplace — plugins — ccplugin-*
3. claude-config (this repo) — full OS: ./install.sh, skills, MCP, hooks, agent-registry.json
4. ClaudeHQ — multi-project workspace hub

Optional: sdk-market. Counts follow main here and the linked repos.

flowchart LR
  A[1 Catalog] --> B[2 Marketplace]
  B --> C[3 claude-config]
  C --> D[4 ClaudeHQ]

A portable, self-installing configuration system that turns Claude Code into a multi-agent operating system. Clone → ./install.sh → done.

• 204 registered agents in agent-registry.json · 15 categories (activate on demand)
• 21 plugins on our marketplace
• 8 MCP servers integrated (GitHub, Atlassian/Jira, Firebase, Flutter, jCodeMunch, Git, Fetch, Context7)
• 54 slash commands (/yolo, /team-build, /jira-run, /audit, /web-research, /sprint-plan, and more)
• Telegram bot with persistent Haiku agent — control Claude from your phone
• Local-first model routing — Ollama → Claude → Groq free (saves tokens)
• Cost control — automatic model tier switching based on quota remaining

git clone https://github.com/SkyWalker2506/claude-config.git ~/Projects/claude-config
cd ~/Projects/claude-config
./install.sh

The installer sets up everything: CLAUDE.md hierarchy, skills, MCP servers, hooks, secrets vault, Telegram integration, and plugin marketplace.

┌─────────────────────────────────────────────────┐
│                  Entry Points                    │
│   Claude Code  ·  Telegram Bot  ·  OpenCode/Zen │
└──────────────────────┬──────────────────────────┘
                       ▼
┌─────────────────────────────────────────────────┐
│              Routing & Budget                    │
│   Task Router → Token Budget → Fallback Manager  │
│           Agent Registry (~204)                  │
└──────────────────────┬──────────────────────────┘
                       ▼
┌───────────┬──────────┬──────────┬───────────────┐
│ Opus (4%) │Sonnet(13%)│Haiku(18%)│ Free/Local(65%)│
└───────────┴──────────┴──────────┴───────────────┘
                       ▲
┌─────────────────────────────────────────────────┐
│              Infrastructure                      │
│  Daily Health Check · Watchdog · Cron Scheduler  │
└─────────────────────────────────────────────────┘

21 plugins, each in its own repo. Add the marketplace:

claude plugin marketplace add SkyWalker2506/claude-marketplace

Then browse with /plugin > Discover or install directly:

Knowledge-First agents across 15 categories (registry on main). Each agent lives in a folder with:

• AGENT.md (identity, boundaries, process)
• knowledge/ (lazy-loaded domain notes)
• memory/ (sessions + learnings)

Production-grade telemetry, lifecycle, and routing across the agent system. All commands write structured reports under Reports/ and config/telemetry/aggregated/.

hq dashboard            # Unified system view (success rate, top agents, model mix)
hq dashboard --json     # Pipeable JSON
hq dashboard --watch    # Live refresh every 60s
hq lifecycle            # Promote/demote/retire recommendations
hq lifecycle --force    # Bypass min sample guard (20 events)
hq optimize             # Cheaper-model swap recommendations from telemetry
hq plugin-eval          # Grade ccplugin-* repos (A–F, 11-point rubric)
hq events               # Tail recent dispatch events
hq stats <agent_id>     # Per-agent breakdown (success, models, projects)
hq health               # Run all of the above in one pass

Telemetry pipeline:

• PostToolUse → log_dispatch.py writes one JSONL line per Task/Bash to config/telemetry/events/<date>.jsonl
• hq_dashboard.py aggregates 30 days into config/telemetry/aggregated/dashboard.json
• agent_lifecycle.py, route_optimizer.py, plugin_eval.py consume the same event stream
• Verification gates (pre-commit.json, pre-push.json) run via gate_check.py PreToolUse hook (warn-only by default)

After install.sh, hq is available on your PATH as ~/Projects/claude-config/bin/hq (or run scripts/hq directly).

Control Claude from your phone — persistent Haiku agent session:

# Start the agent (Haiku, low cost)
python3 config/telegram-agent.py ~/Projects/your-project

# Or classic mode (claude -p --continue per message)
bash config/telegram-poll.sh ~/Projects/your-project

Features: text commands, photo analysis, document processing, voice transcription (Whisper TR/EN), inline keyboard.

~/.claude/CLAUDE.md          ← Global (all projects)
~/Projects/CLAUDE.md         ← Shared rules (framework-agnostic)
~/Projects/MyApp/CLAUDE.md   ← Project-specific

Open claude in any project directory → migration hook fires → setup wizard creates project config.

Fallback priority: Local (Ollama) → Claude (paid) → Free (Groq)

claude-config/
├── install.sh              # One-command setup
├── uninstall.sh            # Restore from backup
├── CLAUDE.md               # Repo-level rules
├── global/
│   ├── CLAUDE.md           # → ~/.claude/CLAUDE.md
│   ├── settings.json.template  # MCP servers, hooks, permissions
│   └── skills/             # 54 slash commands
├── agents/                 # Knowledge-First agent trees
│   ├── orchestrator/       # A1–A8
│   ├── backend/            # B1–B19
│   ├── code-review/        # C1–C6
│   └── ...                 # 15 categories
├── config/
│   ├── agent-registry.json # Agent → model mapping
│   ├── fallback-chains.json
│   ├── model-tiers.json
│   ├── telegram-agent.py   # Persistent Telegram bot
│   └── daily-check.sh      # Health check script
├── hooks/                  # Shared hook scripts → ~/.claude/hooks/
│   ├── session-start-graph.sh        # Emits graph-cache summary on SessionStart
│   └── post-commit-graph-update.sh   # Marks graph-cache stale on git commit
├── plugins/                # Plugin source (telegram, ai-review, etc.)
└── templates/              # Project scaffolding

API keys live in a private git repo, symlinked to ~/.claude/secrets/. Never touches public repos.

# Setup happens during install.sh — your private repo is cloned automatically
# Or manually:
echo 'TELEGRAM_BOT_TOKEN=xxx' >> ~/Projects/claude-config/claude-secrets/secrets.env

git clone https://github.com/SkyWalker2506/claude-config.git ~/Projects/claude-config
cd ~/Projects/claude-config
./install.sh
# Secrets pulled automatically if you own the private repo

• scripts/README.md — audit ve toplu knowledge script’leri (agent_coverage_audit, knowledge_quality_audit, inject_knowledge_maps, vb.).
• Yapısal doğrulama: python3 scripts/verify_knowledge_structure.py — coverage P0–P2 sıfır ve tüm konu dosyalarında mega bölümler tam mı kontrol eder (CI: .github/workflows/knowledge-audit.yml).
• Kaynak derinliği (isteğe bağlı): python3 scripts/knowledge_sources_audit.py — frontmatter sources: istatistiği; çıktı agents/KNOWLEDGE_SOURCES_AUDIT.md.

• AGENT_PLUGIN_MAP.md — navigable index: which plugins work with which agents, which MCPs each active agent requires, and a quick MCP → agents reverse lookup.

• Agent Catalog — agent inventory (GitHub entry point #1)
• Plugin Marketplace — plugins for Claude Code, browse & install
• ClaudeHQ — ecosystem HQ, cross-project workspace
• SDK Market — production-ready SDKs and kits
• OpenCode — Free/local models via Zen + Ollama

Musab Kara — LinkedIn · GitHub

MIT