Agent Tool Hub

One registry, many protocols — MCP, LangChain, n8n, SKILL in one PTC runtime. Source Drop a folder under a root; one tool can expose multiple protocols.

What we support

| Supported tools | How to write | Spec | |-----------------|--------------|------| | SKILL | We fully support the SKILL spec with any LLM.ExamplesSKILL spec and our implementation support | Agent Skills (Anthropic) | | LangChain | Export a StructuredTool in langchain/; we auto-discover.Examples | LangChain Tools | | MCP | Put mcp.json (Cursor-style) in mcp/; we connect as client.We recommend easy-mcp-server for writing MCP servers.Example | MCP Specification | | n8n | Put workflow JSON in n8n/; we run local n8n.@easynet/n8n-local | n8n Workflows |

Install

Node 18+.

Default — MCP / LangChain / SKILL only (~tens of MB):

npm install @easynet/agent-tool-hub

+ n8n — workflows / stock example (~1.3GB):

npm install @easynet/agent-tool-hub @easynet/n8n-local

Run the stock research example

ReAct + yahoo-finance SKILL + HTML report. After install:

npx agent-toolhub-react-stock GOOGL

Ticker: GOOGL, AAPL, MSFT. Set LLM in examples/agent-toolhub-react-stock.mjs or env (OPENAI_API_KEY, OPENAI_BASE_URL). Output: console + GOOGL-research-report.html. Sample report.

| Report | Debug | |--------|-------| | | |

Use

Embed in LangChain

import { createAgentToolHub } from "@easynet/agent-tool-hub/langchain-tools";

// 1. Init runtime (loads tools from toolhub.yaml)
const toolHub = await createAgentToolHub("toolhub.yaml");

// 2. Create your LangChain agent, pass our tools, and run
const agent = createAgent({
  model: new ChatOpenAI({ temperature: 0 }),
  tools: toolHub.tools, // discovered tools from SKILL / LangChain / MCP / n8n
});
const stream = await agent.stream(/* your messages */);
// ...

// 3. Shutdown
await toolHub.shutdown();

Optional: formatStepProgress(step) for console; writeReportFromStream(stream, { htmlReportPath, onStep }) for HTML report.

Code reference

SKILL · LangChain · MCP · n8n.

SKILL

Markdown (SKILL.md) under skill/. Progressive disclosure:

• Level 1 = frontmatter (name, description)
• Level 2 = body (instructions)
• Level 3 = resources (e.g. references/, scripts/, assets/) — scanned from the skill dir, exposed as resource list; agents can reference them by path

Spec & impl.

# skill/SKILL.md
---
name: my-tool
description: What your tool does.
---

# Instructions (Level 2)

Steps the agent should follow when using this skill.

# Level 3 (resources)

Put files under the skill dir (e.g. `references/REFERENCE.md`, `scripts/`, `assets/`). They appear in the resource list and can be read by path.

LangChain

Export a LangChain tool (e.g. StructuredTool). Put under langchain/:

// langchain/calculator.js
import { StructuredTool } from "@langchain/core/tools";
import { z } from "zod";

class CalculatorTool extends StructuredTool {
  name = "calculator";
  description = "Evaluates arithmetic expressions";
  schema = z.object({ expression: z.string() });
  async _call({ expression }) {
    return String(Function(`"use strict"; return (${expression})`)());
  }
}
export default new CalculatorTool();

MCP

MCP client only; put Cursor-style mcp.json under mcp/. MCP servers: easy-mcp-server.

// mcp/mcp.json
{
  "mcpServers": {
    "calculator": {
      "command": "node",
      "args": ["./server.js"]
    }
  }
}

n8n

Workflow JSON under n8n/; local server via @easynet/n8n-local. Optional: npm install @easynet/n8n-local.

// n8n/workflow.json
{
  "name": "My Workflow",
  "nodes": [
    {
      "id": "webhook",
      "type": "n8n-nodes-base.webhook",
      "parameters": { "path": "my-tool", "httpMethod": "POST" }
    }
  ],
  "connections": {}
}