GreenNode AgentBase Skills

A bundle of SKILL.md-compatible skills that drive the full GreenNode AgentBase lifecycle — scaffold → configure → code → test → deploy → monitor → teardown — from inside your AI coding tool.

Drop them into Claude Code, Cursor, OpenAI Codex, or any other SKILL.md-aware client and you get slash commands like /agentbase-wizard, /agentbase-deploy, /agentbase-monitor. The skills are plain Markdown + shell — no client-specific runtime — so the full lifecycle works in every tool that can read SKILL.md and run a shell.

---

TL;DR — Install in 30 Seconds

git clone https://github.com/vngcloud/greennode-agentbase-skills.git

# Pick the install target for your tool (see table below)
#   Claude Code  → ~/.claude/skills        or  <project>/.claude/skills
#   Cursor       → ~/.cursor/skills        or  <project>/.cursor/skills
#   Codex        → ~/.agents/skills        or  <project>/.agents/skills

mkdir -p ~/.claude/skills
cp -r greennode-agentbase-skills/.claude/skills/* ~/.claude/skills/

Then restart your tool and type /agentbase-wizard (or just say "build me a Telegram bot").

---

Install Per Tool

All skills live under .claude/skills/. They are plain folders with a SKILL.md file inside — no build step. Each client auto-discovers them from a known directory.

1. Claude Code

The native home for these skills.

# Global (recommended — available in every project)
mkdir -p ~/.claude/skills
cp -r greennode-agentbase-skills/.claude/skills/* ~/.claude/skills/

# OR project-scoped
mkdir -p <your-project>/.claude/skills
cp -r greennode-agentbase-skills/.claude/skills/* <your-project>/.claude/skills/

Launch and use:

cd <your-project> && claude
> /agentbase-wizard          # slash command
> "deploy my agent"           # or just describe intent — Claude picks the skill

Tip: claude will auto-load every SKILL.md it finds. To verify, run /help and look for the skills section.

2. Cursor

Cursor's skills support and exact path have evolved across releases — check your version's docs for the correct skills directory before installing. Typical layout:

mkdir -p ~/.cursor/skills
cp -r greennode-agentbase-skills/.claude/skills/* ~/.cursor/skills/
# or project-scoped: <your-project>/.cursor/skills/

Open Cursor → Agent chat → type / to search skills. Agent mode runs bash / curl, so deploy / monitor / teardown work end-to-end.

3. OpenAI Codex

export OPENAI_API_KEY="..."
cd <your-project> && codex

Codex CLI reads SKILL.md-style files; the exact discovery path depends on your Codex version (commonly ~/.agents/skills/ or <project>/.agents/skills/ — check your version's docs). Once discovered, the CLI executes shell + HTTP calls, so the full lifecycle works.

4. Other SKILL.md-compatible Clients

Any client that (a) reads SKILL.md frontmatter (name, description) and (b) can run shell commands will work. Point the client at the .claude/skills/ directory or copy folders into whatever skills path it expects.

Compatibility Matrix

The skills are tool-agnostic — they're just Markdown procedures plus bash / curl calls to the GreenNode REST APIs. Every SKILL.md-aware client with shell access can run them end-to-end. Differences below are about UX, not capability.

	Claude Code	Cursor	Codex	Other SKILL.md clients
Typical skills directory	.claude/skills/	.cursor/skills/	.agents/skills/	client-specific
Invocation	/skill-name	/skill-name (Agent)	natural language / CLI	varies
Auto-routing by description	✅ native	✅	✅	depends on client
Runs shell / HTTP from skills	✅	✅ (Agent mode)	✅	requires shell tool
Full deploy & monitor pipeline	✅	✅	✅	✅ if shell available

The skills are authored and tuned primarily on Claude Code — that's where routing and prompts are validated. Functionally though, every tool with shell access can run them; Cursor / Codex / other clients just don't have a dedicated test pass yet.

---

Prerequisites

Before any skill that hits the platform, set GreenNode IAM credentials:

export GREENNODE_CLIENT_ID="<service-account-client-id>"
export GREENNODE_CLIENT_SECRET="<service-account-secret>"

Put them in your shell profile or in a project-local .env (never commit it — .env.example is the tracked template).

Skills that only read local files (e.g. agentbase-wizard init) work without credentials.

---

Skills Index

Skill	What it does
/agentbase-wizard	Start here. Guided 9-step lifecycle: scaffold → configure → code → test → deploy → verify. Also handles init, test, resume.
/agentbase	Platform reference — architecture, services, IAM, "which skill should I use".
/agentbase-identity	Register agent identities; store API keys / OAuth2 credentials for external services (OpenAI, Google, Slack, …).
/agentbase-llm	Manage platform LLM access — API keys, model catalog, rate limits, OpenAI-compatible endpoint.
/agentbase-memory	Conversation history, semantic memory, long-term memory stores (LangChain/LangGraph integration).
/agentbase-deploy	Build & push Docker image, create/update Custom Agent runtimes (PUBLIC/VPC), deploy OpenClaw Telegram/Zalo bots, manage the Container Registry.
/agentbase-monitor	Runtime logs, endpoint logs, CPU/RAM metrics, unified resource dashboard.
/agentbase-gateway	Resource Gateway (MCP) CRUD; inbound auth (NONE / IAM / JWT); per-target outbound auth (APIKEY / OAUTH 2LO / 3LO); VPC routes; Policy Group binding.
/agentbase-policy	Authorization policies — Policy Groups, Policies, and statement bodies (effect / principal / actions / resources / condition). Enforced today on the Resource Gateway.
/agentbase-teardown	Delete all resources for a project. Always supports --dry-run.

Lifecycle Map

┌────────────────────────────────────────────────────────┐
│ GET STARTED                                            │
│   /agentbase-wizard ────── guided A → Z                │
│   /agentbase ───────────── platform reference          │
├────────────────────────────────────────────────────────┤
│ BUILD & CONFIGURE                                      │
│   /agentbase-wizard init ── scaffold project           │
│   /agentbase-llm ────────── platform LLM access        │
│   /agentbase-identity ───── identities & external auth │
│   /agentbase-memory ─────── memory stores              │
├────────────────────────────────────────────────────────┤
│ TEST & DEPLOY                                          │
│   /agentbase-wizard test ── validate / local / docker  │
│   /agentbase-deploy ─────── build, push, deploy        │
├────────────────────────────────────────────────────────┤
│ OPERATE                                                │
│   /agentbase-monitor ────── logs, metrics, dashboard   │
│   /agentbase-gateway ────── Resource Gateway (MCP)     │
│   /agentbase-policy ─────── access policies            │
├────────────────────────────────────────────────────────┤
│ ADVANCED                                               │
│   /agentbase-deploy cr ──── Container Registry         │
│   /agentbase-teardown ───── delete everything          │
└────────────────────────────────────────────────────────┘

Common Subcommands

/agentbase-wizard   [init <name> [--langchain|--langgraph] | test [validate|local|docker|preflight] | resume | step-N | reset]
/agentbase-identity identity <create|list|get|update|delete>          [name]
                    auth     <apikey|delegated|oauth2> <create|list|get|update|delete|retrieve> [name]
/agentbase-llm      <api-keys|models> <create|list|get|update|delete|enable|disable|rate-limit> [name-or-uuid]
/agentbase-memory   memory  <create|list|get|delete> [id]
                    events  <list|create|delete>
                    records <browse|search|generate-from-session|generate-from-content|insert|delete>
/agentbase-deploy   Custom Agent: build → push → deploy, runtime CRUD, scale, versions
                    OpenClaw:     create|list|start|stop|switch-version (Telegram/Zalo templates)
                    Container Registry: repo info, credentials, images, artifacts
/agentbase-monitor  <runtime-logs|endpoint-logs|metrics|dashboard> [runtime-id] [endpoint-id]
/agentbase-gateway  <create|list|get|update|delete|routes|repair|flavors> [gateway-name]
/agentbase-policy   <group|policy> <create|list|get|update|delete> [group-id-or-name] [policy-id-or-name]
/agentbase-teardown <project-name> [--dry-run]

These skills are driven by natural language — the syntax above is a quick reference, not a strict CLI. Tell the model what you want and it picks the right operation.

---

End-to-End Example — Build a Chatbot

/agentbase-wizard init my-chatbot --langgraph   # scaffold
/agentbase-llm api-keys create my-chatbot-key   # platform LLM key
/agentbase-memory create                         # optional memory store
/agentbase-wizard test local                     # smoke test locally
/agentbase-deploy deploy                         # build → push → deploy
/agentbase-monitor runtime-logs <runtime-id>     # watch it run

Or, first time, just:

/agentbase-wizard

…and follow the prompts.

---

Troubleshooting

Symptom	Fix
Skill doesn't appear	Confirm the file is at <skills-dir>/<skill-name>/SKILL.md with valid name + description frontmatter, then restart the tool.
401 Unauthorized	GREENNODE_CLIENT_ID / GREENNODE_CLIENT_SECRET missing, expired, or service account lacks IAM policies.
OOMKilled during deploy	Pick a larger flavor — ask /agentbase-deploy to list eligible flavors and resize the runtime.
Want to resume a half-finished session	State persists in .agentbase-state.json — run /agentbase-wizard resume.
Different skill behavior across tools	Tightest validator wins (typically Claude Desktop). Re-read the SKILL.md frontmatter; description length / characters may need trimming.

---

Repo Layout

greennode-agentbase-skills/
├── .claude/skills/             # <-- the skills you install
│   ├── agentbase/              # platform reference
│   ├── agentbase-wizard/       # guided full-lifecycle wizard
│   ├── agentbase-deploy/       # build, push, deploy + Container Registry + OpenClaw
│   ├── agentbase-identity/     # agent identities & outbound auth
│   ├── agentbase-llm/          # platform LLM API keys & models
│   ├── agentbase-memory/       # conversation + semantic memory
│   ├── agentbase-monitor/      # logs, metrics, dashboard
│   ├── agentbase-gateway/      # Resource Gateway (MCP)
│   ├── agentbase-policy/       # authorization policies
│   └── agentbase-teardown/     # delete all resources
└── README.md

Each skill folder contains a SKILL.md (the contract read by the AI tool) and any helper scripts/ or references/ it needs.

---

Contributing & Extending

• Each skill is just a folder with a SKILL.md file — copy an existing one as a template.
• Frontmatter (name, description) is a public contract — renaming breaks downstream tools. Update README.md cross-references when you rename.
• Skill descriptions must satisfy the tightest client validator (typically Claude Desktop's character limit). Verify before committing.
• Test the skill end-to-end in Claude Code (or your target client) before opening a PR — descriptions drive auto-routing, so a small wording change can shift which skill is picked.

---

Important Notes

1. Verify IAM credentials first — the majority of platform errors are missing GREENNODE_CLIENT_ID / GREENNODE_CLIENT_SECRET or insufficient policies.
2. Validate before deploy — /agentbase-wizard test validate.
3. Always --dry-run teardown before the real delete.
4. Never commit .env — only .env.example is tracked.
5. First time? Use /agentbase-wizard — it covers the full 9-step path.