🤖 n8n-nodes-agent-kit

⚡ Why Agent Kit?

n8n's built-in AI Agent is great for prototypes. Agent Kit is what you reach for when you ship to production.

One package. Six nodes. Zero magic. Plug, wire, ship.

🚀 Quick Start

1. Install

In your n8n instance: Settings → Community Nodes → Install → paste:

n8n-nodes-agent-kit

Or via npm in self-hosted:

npm install n8n-nodes-agent-kit

2. Add credentials

• OpenRouter API — grab one at openrouter.ai (gives you Claude, GPT-4o, Gemini, Llama, Qwen, etc. — one key)
• GitHub Skills API (optional) — only if you want to load skills from a repo

3. Build your first agent (60 seconds)

[Webhook] → [Agent Memory] ─┐
                            ├→ [Agent Kit] → [Respond]
            [MCP Gateway] ──┘

Done. You have a stateful AI agent with tools.

🧩 The Nodes

🎬 Examples

🟢 Single agent with memory + tools

[Trigger] → [Agent Memory] ─┐
                            ├→ [Agent Kit] → [Output]
            [MCP Gateway] ──┘

🔵 Multi-agent orchestration

                  ┌→ [SubAgent: researcher] ──┐
[Trigger] → [Orchestrator Kit] ←──────────────┤
                  └→ [SubAgent: writer]    ───┘

The orchestrator picks which subagent to call. Each subagent has its own prompt, tools, and (optionally) memory. The output includes executionTrace with per-agent token usage.

🟣 Skills from GitHub (agentskills.io style)

[Trigger] → [Skill Loader (GitHub)] → [Agent Kit]

Skill Loader fetches SKILL.md files from your repo, parses YAML frontmatter, and injects them into the agent's system prompt. Update your skills repo, no n8n redeploy.

A complete demo workflow lives in docs/ — import it and you're running in 30 seconds.

🛡️ Guardrails

Agent Kit and Orchestrator Kit ship with built-in input/output guardrails:

• PII detection (email, phone, CPF, etc.)
• Regex blocklists
• Length caps
• Custom blockers

Configure them per-node. No middleware. No proxy. Just toggle.

📊 Observability

Every Orchestrator run returns:

{
  "response": "...",
  "executionTrace": [
    { "step": 1, "agent": "researcher", "task": "...", "durationMs": 1840,
      "usage": { "prompt_tokens": 1230, "completion_tokens": 240, "iterations": 2 } },
    { "step": 2, "agent": "writer", "task": "...", "durationMs": 920,
      "usage": { "prompt_tokens": 800, "completion_tokens": 410, "iterations": 1 } }
  ]
}

Wire this straight into your dashboards. No extra instrumentation.

🛠️ Local Development

git clone https://github.com/lucasbrito-wdt/n8n-nodes-agent-kit.git
cd n8n-nodes-agent-kit
npm install
npm run build
npm link

# In your n8n install dir:
npm link n8n-nodes-agent-kit
n8n start

Run the test suite:

npm test

🤝 Contributing

This is the part where you come in.

Agent Kit is built in the open. PRs, issues, ideas, memes — all welcome.

• 🐛 Report a bug
• 💡 Request a feature
• 📖 Read the Contributing Guide
• 🧭 Follow the Code of Conduct

First-time contributors welcome. Look for issues tagged good first issue.

🗺️ Roadmap

• Multi-agent orchestration (Orchestrator + SubAgent)
• Per-agent token usage in executionTrace
• MCP client + server in one node
• Skills from GitHub
• Streaming responses
• Agent eval harness (golden datasets + regression gates)
• Built-in vector memory
• Cost budgets & per-session caps
• Web UI for inspecting executionTrace

Vote with 👍 reactions on the roadmap discussion.

📜 License

MIT © Lucas Brito