🤖 n8n-nodes-agent-kit

Build production-grade AI agents inside n8n. No glue code. No black boxes.

Multi-agent orchestration · Persistent memory · MCP tools · Skills as Code · Guardrails — all native n8n nodes.

Quick Start · Nodes · Examples · Contributing · Discord

⚡ 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.

| | Vanilla n8n AI Agent | Agent Kit | |---|:---:|:---:| | Multi-agent orchestration (supervisor → subagents) | ❌ | ✅ | | Persistent SQLite long-term memory | ❌ | ✅ | | Skills loaded from GitHub (agentskills.io) | ❌ | ✅ | | MCP client + server gateway in one node | ❌ | ✅ | | Per-agent token usage trace | ❌ | ✅ | | Built-in guardrails (PII / regex / length) | ❌ | ✅ | | Drop-in OpenRouter (400+ models) | ⚠️ | ✅ |

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

🤖 Agent Kit

The workhorse. A single AI agent with tool calling, memory, skills, and guardrails. Use this when one agent is enough.

Inputs: Main · AiMemory · AiTool Powered by: any OpenRouter model

🧠 Orchestrator Kit

Supervisor agent that routes tasks to specialist subagents. Returns a full executionTrace showing who did what, when, and how many tokens.

Inputs: Main · AiMemory · AiTool · AiAgent

🎯 Sub Agent Kit

A specialized agent that plugs into an Orchestrator as a callable tool. Give it a name, a description, a prompt — done. The orchestrator decides when to call it.

Output: AiAgent

💾 Agent Memory

SQLite-backed long-term memory + in-process short-term session buffer. Survives restarts. No external DB required.

Output: AiMemory

🔌 MCP Gateway

Speaks Model Context Protocol both ways: client (consume external MCP servers as tools) and server (expose n8n webhooks as MCP tools).

Output: AiTool

📚 Skill Loader

Loads SKILL.md files (the agentskills.io format) inline or straight from a GitHub repo. Skills compose into the system prompt at runtime.

Output: Main (carries skills payload)

🎬 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

• [x] Multi-agent orchestration (Orchestrator + SubAgent)
• [x] Per-agent token usage in executionTrace
• [x] MCP client + server in one node
• [x] 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

If Agent Kit saved you a week of glue code, drop a ⭐ — it's the cheapest thank-you in OSS.

⭐ Star on GitHub · 📦 npm · 🐛 Issues

Built with ❤️ for the n8n community.