Zenox for Claude Code

Multi-agent orchestrator plugin for Claude Code. Gives you a team of specialized AI agents instead of a single generalist.

What it does

Zenox replaces the default Claude Code agent with an orchestrator that automatically delegates to specialists:

| Agent | Model | Role | |-------|-------|------| | Zenox (main) | Opus | Orchestrates everything, writes code, delegates when needed | | Explorer | Haiku | Fast codebase search — finds files, patterns, symbols | | Librarian | Sonnet | External research — official docs, GitHub examples, library internals | | Oracle | Opus | Strategic advisor — architecture decisions, debugging, code review |

Explorer and Librarian run in background automatically. Oracle runs in foreground (blocks until result arrives). You don't manage the agents — the orchestrator delegates based on what you ask.

Install

From marketplace (recommended)

/plugin marketplace add CyberBoyAyush/zenox-claude
/plugin install zenox@zenox-marketplace
/reload-plugins

From npm

If the marketplace is already configured to use npm:

/plugin marketplace add CyberBoyAyush/zenox-claude
/plugin install zenox@zenox-marketplace

Local development

claude --plugin-dir /path/to/zenox-claude

How delegation works

You don't call agents directly. The main agent delegates automatically:

You: "Where's the authentication logic?"
-> Explorer searches the codebase (background)

You: "How does React Query handle caching?"
-> Librarian fetches docs + real GitHub examples (background)

You: "Should I use Redux or Zustand here?"
-> Oracle analyzes trade-offs (foreground)

You: "review"
-> Oracle reviews your recent git diff for bugs, security, regressions

Skills (keyword triggers)

| Skill | Trigger | What happens | |-------|---------|--------------| | /ultrawork | "ultrawork" or "ulw" | Maximum multi-agent coordination — fires parallel agents | | /deep-research | "deep research" | 3-4 agents explore internal + external sources | | /explore-codebase | "explore codebase" | Multiple explorers map out relevant code | | /review | "review", "self-review" | Oracle reviews your git diff |

Included MCP servers

These are auto-loaded when the plugin is enabled:

| Server | Purpose | |--------|---------| | exa | Web search, docs lookup, URL crawling (free, no API key needed) | | grep-app | Search millions of public GitHub repos | | sequential-thinking | Step-by-step reasoning for complex problems |

All three servers are publicly accessible and require no API keys or configuration.

Self-review protocol

After significant implementations (3+ files changed, security-sensitive code, architecture changes), the orchestrator automatically:

1. Runs git diff to capture changes
2. Fires Oracle with the complete diff
3. Collects the review verdict
4. Addresses Critical/High findings before delivering

Skipped for trivial changes (single file, <50 lines, docs/config only).

Plugin structure

zenox-claude/
├── .claude-plugin/
│   ├── plugin.json          # Plugin manifest
│   └── marketplace.json     # Marketplace catalog
├── agents/
│   ├── zenox.md             # Main orchestrator agent
│   ├── explorer.md          # Codebase search specialist
│   ├── librarian.md         # External research specialist
│   └── oracle.md            # Strategic advisor + code reviewer
├── skills/
│   ├── ultrawork/SKILL.md
│   ├── deep-research/SKILL.md
│   ├── explore-codebase/SKILL.md
│   └── review/SKILL.md
├── hooks/
│   └── hooks.json
├── .mcp.json                # MCP server configs (exa, grep-app, sequential-thinking)
├── settings.json            # Activates zenox as main agent
└── package.json             # npm distribution

Configuration

The plugin activates the zenox agent as the main thread via settings.json. Model assignments are fixed to Claude's available tiers:

• Explorer: Haiku (fast, cheap — right for grep-like tasks)
• Librarian: Sonnet (balanced — good for research synthesis)
• Oracle: Opus (maximum reasoning — right for architecture/review)
• Zenox: Opus (main thread needs full capability)

Claude Code only supports Anthropic models (haiku/sonnet/opus) for agents, so there's no model configuration step.

Differences from Zenox for OpenCode

This is a port of Zenox (OpenCode plugin) adapted for Claude Code's declarative plugin system:

| Aspect | OpenCode Zenox | Claude Code Zenox | |--------|---------------|-------------------| | Implementation | TypeScript plugin (~1000 LOC) | Static markdown files | | Agent count | 4 (includes UI Planner) | 3 (no UI Planner) | | Model flexibility | Any provider (OpenAI, Google, Anthropic) | Anthropic only | | Background tasks | Custom BackgroundManager | Native background: true frontmatter | | Model config | Interactive CLI (bunx zenox config) | Fixed in agent frontmatter | | Distribution | npm + OpenCode plugin registry | npm + Claude Code marketplace |

License

MIT