A 12-agent specialist config pack for Claude Code.
Twelve color-coded agents, each owning exactly one concern, with zero overlap.
A single coding assistant tries to do everything at once: architecture, UX, performance, security, cleanup. The concerns blur together and things slip through. Nexus splits the work into 12 named specialists, each with a fixed scope. When Claude Code touches your code, the relevant agents review their slice and nothing else. One orchestrator, Snow, assigns work and gives final approval, so reviews stay structured instead of sprawling.
- 12 specialist agents, each owning one concern with no overlap
- Color-coded roles for instant recognition (white orchestrator, red architecture, and so on)
- Ships in four ready-to-use formats:
CLAUDE.md, a/nexusskill,agents.json, and a React component - A fixed report format so every agent answers in the same shape
- A single orchestrator (Snow) that assigns tasks and owns the final approval gate
- Drop-in, dependency-free, framework-agnostic
| # | Agent | Role | Color | Scope |
|---|---|---|---|---|
| 1 | Snow | Orchestrator | White | Final approval, task assignment, full visibility |
| 2 | Blaze | Architecture | Red | System design, module boundaries, PR reviews |
| 3 | Arrow | UX / QA | Pink | E2E tests, accessibility, responsiveness |
| 4 | Venus | UI / Branding | Orange | Colors, typography, spacing, brand consistency |
| 5 | Zap | Performance | Yellow | Re-renders, bundles, Lighthouse, CLS, query efficiency |
| 6 | Frost | Metrics | Cyan | Lighthouse tracking, bundle analysis, quality trends |
| 7 | Blitz | Code Quality | Blue | Unit tests, coverage, lint, type safety |
| 8 | Earth | Cleanup | Green | Dead code, deduplication, file organization |
| 9 | Pulse | Automation | Purple | CI pipelines, build gating, dev tooling |
| 10 | Sand | Storage | Brown | localStorage, Supabase, caching, migrations |
| 11 | Shadow | Security | Grey | Auth audits, XSS, header hardening, API security |
| 12 | Rock | Dashboard | Slate | Admin views, data tables, monitoring UI |
Clone the repo and copy whichever format you want into your project:
git clone https://github.com/bunlongheng/nexus| File | Use it as |
|---|---|
CLAUDE.md |
Project instructions Claude Code reads automatically |
.claude/commands/nexus.md |
A /nexus slash command skill |
agents.json |
Machine-readable config for tooling, dashboards, or CI |
components/AgentWiki.tsx |
An interactive React wiki for a Next.js app |
As CLAUDE.md instructions. Copy CLAUDE.md into your project root. Claude Code reads
it on every session and applies the agent roles automatically.
As a Claude Code skill. Copy .claude/commands/nexus.md into your project, then run
/nexus in Claude Code to load the 12-agent system for that session.
As a JSON config. Import agents.json into your own tooling, dashboards, or CI
pipelines. It holds each agent's id, name, role, color, scope, and description.
As a React component. Drop components/AgentWiki.tsx into a Next.js app for an
interactive, animated agent wiki page.
Each agent has a dedicated scope, so reviews never collide:
- Editing a component? Venus checks branding, Arrow checks UX, Blitz checks types.
- Adding an API route? Shadow audits security, Sand checks queries, Zap checks performance.
- Deploying? Pulse verifies CI, Frost tracks metrics, Snow gives final approval.
No agent steps on another's territory, and Snow orchestrates the whole flow. Every agent
answers in the same fixed report format: an H1 title, ALL-CAPS section headers, dash
bullets, and a closing SUMMARY of 3 to 6 bullets, so output stays scannable no matter
which specialist produced it.