Stop prompting. Start operating.
Turn Claude Code & OpenAI Codex into structured AI coding agents

→ 🧠 20 specialized agents (backend, security, UI, etc.)
→ 🧩 42 modular skills (loaded on demand)
→ ⚙️ 17 workflows (brainstorm → plan → build → test → deploy)

AgentHub gives your AI coding sessions an instant upgrade: domain experts that activate automatically, deep knowledge modules loaded on demand, and workflow commands that orchestrate entire feature builds — all on your budget, with your consent.

Platform difference: Claude Code uses /hub: slash commands. Codex uses @hub mentions. The / prefix is reserved for Codex built-ins — typing /hub:debug in Codex will be rejected. Always use @hub debug in Codex.

tools/generate_codex.py keeps the Codex agent TOML files in sync with the Claude source on every release.

Claude Code and Codex ship powerful runtimes, but every project reinvents the same scaffolding:

• Who is the "backend expert" and what do they know?
• Which skills load for a database change vs. a UI change?
• How do you orchestrate multiple agents without burning your token budget?

AgentHub gives you a curated, MIT-licensed answer. One install, both ecosystems.

VS Code / Cursor extension:

1. Press / in the Claude Code sidebar → Manage Plugins → Marketplace
2. Paste https://github.com/SKB3002/agenthub and confirm
3. Go to Plugins, search hub, and click Install
4. Press Ctrl+Shift+P -> Reload Window

Run /hub:help to verify — it will list all 17 commands, 20 agents, and 42 skills or press '/' and manually check new loaded commands

# 1. Install the Codex CLI (if you haven't already)
npm install -g @openai/codex

# 2. Register the AgentHub marketplace (either command works)
codex plugin marketplace add https://github.com/SKB3002/agenthub
# or the older short form:
codex marketplace add https://github.com/SKB3002/agenthub

# 3. Start Codex in PowerShell (or your terminal)
codex

Inside Codex: type /plugin → search hub → Install.

Use @hub help to verify. Always use @hub in Codex, never /hub: — the / prefix is reserved for Codex built-ins.

1. /hub:brainstorm <idea>        — explore options before writing any code  (MEDIUM)
2. /hub:plan <feature>           — turn the chosen direction into a plan file  (MEDIUM)
3. /hub:create <app>             — scaffold a greenfield app  (HEAVY)
   or /hub:enhance <change>      — add/update features in an existing app  (HEAVY)
4. /hub:test generate <module>   — generate tests for the new code  (MEDIUM)
   or /hub:test                  — run existing tests  (LIGHT)
5. /hub:deploy staging           — pre-flight + deploy  (HEAVY)
   then /hub:deploy production

Sprinting on a bug? /hub:debug → /hub:test → done.

1. @hub brainstorm <idea>        — explore options before writing any code
2. @hub plan <feature>           — turn the chosen direction into a plan file
3. @hub create <app>             — scaffold a greenfield app
   or @hub enhance <change>      — add/update features in an existing app
4. @hub test generate <module>   — generate tests for the new code
   or @hub test                  — run existing tests
5. @hub deploy staging           — pre-flight + deploy
   then @hub deploy production

Sprinting on a bug? @hub debug → @hub test → done.

Every MEDIUM/HEAVY command renders an approval gate before dispatching agents — you always see what will run and can cancel or pick a lighter alternative. Pass --yes / -y to bypass.

Use /hub:<command> in Claude Code · Use @hub <command> in OpenAI Codex

Dispatch via Agent(subagent_type="hub:<name>"). All agents are namespaced hub:.

Skills are loaded on demand by agents. All 42 live in skills/ and are shared across platforms:

• Cross-stack quality: hub:clean-code, hub:lint-and-validate, hub:testing-patterns, hub:tdd-workflow, hub:systematic-debugging, hub:code-review-checklist
• Backend / Python: hub:fastapi-expert, hub:sqlalchemy-expert, hub:python-patterns, hub:api-patterns, hub:database-design
• LLM / AI: hub:llm-observability, hub:mcp-builder
• Frontend: hub:frontend-design, hub:web-design-guidelines, hub:tailwind-patterns, hub:nextjs-react-expert, hub:mobile-design
• Workflow: hub:socratic-gate, hub:approval-gate, hub:plan-writing, hub:parallel-agents, hub:intelligent-routing, hub:behavioral-modes
• Security: hub:vulnerability-scanner, hub:red-team-tactics
• Ops: hub:deployment-procedures, hub:server-management, hub:performance-profiling

Full list: skills/. Run /hub:help skills (Claude Code) or @hub help skills (Codex) to list everything live.

Every command declares a tier. Commands never silently fan out agents on your tokens.

Pass --yes or -y as the first argument to bypass any gate.

Usage log — .hub/usage.json (gitignored, project-local)

Read it via /hub:ledger (Claude Code) or @hub ledger (Codex):

• ledger weekly — ISO-week totals by tier, top agents, top skills
• ledger by-agent / by-skill / by-tier — ranked aggregations
• ledger roi — useful/wasted/partial ratios (from runs you tag)

Honesty contract: every token number is prefixed ~. No dollar conversion, ever.

claude/hooks/hooks.json ships empty. To wire lint-on-edit, security-on-save, or test-on-stop, copy entries from claude/hooks/hooks.example.json into claude/hooks/hooks.json and restart Claude Code.

Copy entries from claude/mcp.example.json into your project's .mcp.json to activate any of the 5 pre-validated servers (filesystem, github, postgres, fetch, sequential-thinking).

PRs welcome. See CONTRIBUTING.md. Whether you're adding a subagent, fixing a skill, or proposing a workflow — all contributions are appreciated.

• This Repo is Built based on the antigravity-kit by @vudovn — the structural blueprint. Licensed MIT.
• Skill content draws from Vercel's React & Web Design Guidelines, OWASP, Core Web Vitals, and the broader open-source community.

MIT — see LICENSE. Use it commercially, personally, or in your side project. Forks, remixes, and derivatives are welcome.