Vincent

Posted on Mar 5

Building an AI Agent? Stop Rebuilding These 13 Utility Tools Every Time

#javascript #ai #webdev #mcp

Every time I start a new AI agent project, I hit the same wall.

The LLM reasoning part? That's the fun bit.

The annoying part is wiring up the same handful of utilities I always end up needing:

Web search to get current information
URL content extraction (clean text for LLM context)
Screenshot capabilities for visual tasks
Weather data
Stock prices and FX rates
Email validation before sending outreach
Translation for global reach
News feed access
WHOIS + DNS lookups
GeoIP resolution
PDF extraction
QR code generation

Each one has its own credentials, rate limit quirks, error formats, and output that needs massaging before it's useful in an LLM context. I've rebuilt this stack more times than I care to admit.

So I extracted it all into a standalone service: Agent Toolbox — an open-source MCP server + REST API that bundles 13 production-ready tools under one key.

What's included

Tool	Description	Under the hood
`search`	Structured web search, paginated results	DuckDuckGo via ddg-node
`extract`	Clean content from any URL, strips boilerplate	Mozilla Readability
`screenshot`	Full-page PNG screenshots	Playwright/Chromium
`weather`	Current conditions + 7-day forecast	Open-Meteo (no key needed)
`finance`	Stock quotes, historical data, FX rates	Yahoo Finance
`validate-email`	Format + DNS MX + SMTP probe	Custom implementation
`translate`	Multi-language text translation	Cloud translation backend
`news`	Latest headlines by category or keyword	News aggregation
`whois`	Domain registration + expiry info	WHOIS protocol
`dns`	DNS record lookup (A, MX, TXT, CNAME, etc.)	DNS resolver
`geoip`	IP geolocation with country/city/ASN	GeoIP database
`pdf_extract`	Extract clean text from PDF URLs	PDF parsing
`qr_generate`	Generate QR codes as PNG or SVG	QR library

Free tier: 1,000 calls/month, no credit card.

Why bundle instead of calling APIs directly?

Fair question. Here's what I actually found building agents:

1. Uniform error handling

Each underlying API fails differently. Rate limits, network timeouts, parsing failures — Agent Toolbox normalizes all of these into consistent error responses. Your agent handles one error format, not thirteen.

2. Single auth token

Instead of injecting 7–13 different API keys into your agent's system prompt, you pass one. Less credential management, less context waste.

3. LLM-optimized output

Raw API responses are noisy for LLMs. Agent Toolbox normalizes, truncates, and formats output for LLM consumption. The extract tool, for example, strips navigation, ads, and boilerplate — returning only article content in clean markdown or plain text.

4. MCP-native

Works out of the box with Claude Desktop, Cursor, Cline, and any MCP-compatible client. Zero custom integration code.

Getting started in 30 seconds

Option 1: Claude Desktop

Add to your claude_desktop_config.json:

{
  "mcpServers": {
    "agent-toolbox": {
      "command": "npx",
      "args": ["-y", "agent-toolbox-mcp"],
      "env": {
        "AGENT_TOOLBOX_API_KEY": "your-key"
      }
    }
  }
}

Restart Claude Desktop. Done. All 13 tools are now available to Claude.

Option 2: Smithery (one click)

Search agent-toolbox on Smithery. Smithery handles the config file automatically.

Option 3: Glama

Also listed on Glama.

Option 4: REST API directly

# Web search
curl -X POST https://api.agenttoolbox.dev/search \
  -H "Authorization: Bearer YOUR_KEY" \
  -H "Content-Type: application/json" \
  -d '{"query": "latest MCP servers 2026", "count": 5}'

# Extract clean content from a URL
curl -X POST https://api.agenttoolbox.dev/extract \
  -H "Authorization: Bearer YOUR_KEY" \
  -H "Content-Type: application/json" \
  -d '{"url": "https://example.com/article", "format": "markdown"}'

Full API docs at Agent Toolbox on GitHub.

Option 5: Python (LangChain / LlamaIndex)

Already on PyPI — two wrappers available:

pip install agent-toolbox-langchain
pip install agent-toolbox-llamaindex

Real-world use cases

Deep research agent

search("AI agent frameworks 2026") → 10 URLs
extract(url) for each → clean article text
LLM synthesizes → structured report

No HTML soup. No rate limit juggling across 3 APIs. Just results.

Competitive intelligence

screenshot + extract competitor landing pages on a schedule. Diff the changes over time. Your agent watches competitors while you sleep.

Financial data pipeline

finance gives real-time stock quotes and FX rates without a Bloomberg subscription. Useful for trading signal agents, portfolio trackers, market commentary.

Email outreach agent

Before sending 500 emails: validate-email checks format, DNS MX records, and SMTP reachability. Protects sender reputation before the campaign fires.

Domain research

whois + dns together let you research domain availability, ownership history, and infrastructure setup — useful for competitive research agents.

Architecture

The server runs on:

Node.js + Hono — lightweight, fast HTTP framework
TypeScript — fully typed throughout
PM2 — process management on VPS
MCP stdio transport — for Claude Desktop / Cursor integration

The MCP layer is a thin wrapper around the REST API. Same tool logic, two interfaces.

Claude Desktop → MCP stdio → Agent Toolbox server → tool execution
Your agent → REST API → Agent Toolbox server → tool execution

MIT licensed. The full source is on GitHub.

Roadmap

Active development. Things coming next:

code-run — sandboxed code execution (most requested)
calendar — timezone-aware date/time helpers
Streaming responses for long extractions
Webhook callbacks for async tool results

Try it

📦 npm MCP package: npx agent-toolbox-mcp
🐙 GitHub (MIT): https://github.com/Vincentwei1021/agent-toolbox
🔧 Smithery: https://smithery.ai/server/@vincentwei1021/agent-toolbox
🔧 Glama: https://glama.ai/mcp/servers/@Vincentwei1021/agent-toolbox
🐍 PyPI: pip install agent-toolbox-langchain

Free tier: 1,000 calls/month. No credit card. If you build something with it, I'd love to hear about it in the comments.

I'm running a $0 multi-agent AI startup experiment. All four agents (CEO, marketer, builder, scout) are live and autonomous. The full story thread is on Twitter/X.

DEV Community