Wide research, not deep reports

Ground your AI agents in current reality.

---

Wide Research

12 sources in 40 seconds. Not one deep dive.

Daily coding is hundreds of micro-decisions: debug this error, check if that library is maintained, find the right API pattern. Deep research tools give you thorough reports—but you don't need a 10-minute analysis when you're mid-task.

Wide research synthesizes many sources fast. Think: asking 12 developers for quick takes vs. hiring one consultant for a report.

Task	Deep Research	Wide Research
"Best auth library for Next.js 15?"	10-min report from 3 articles	40-sec synthesis from 12+ sources
"Why is my React hydration failing?"	Detailed explanation you'll skim	Actual solutions from real discussions
"Is passport.js still maintained?"	History and full analysis	Current status + alternatives if not

When to use each:

• Wide (95% of coding): Debugging, API lookups, library choices, error messages, compatibility checks
• Deep (5%): Learning new frameworks, architecture decisions, strategic planning

---

The Problem

LLMs guess from outdated training data. When your agent needs current info—recent CVEs, breaking changes, deprecation notices—it's working from stale patterns instead of facts.

The Solution

ArguSeek is an MCP server that gives AI agents real-time web research:

• 12+ sources in ~40 seconds — not one deep dive
• Bias detection — promotional content flagged with counter-queries
• Context chaining — build on previous queries without repeating work

The Tools

research_iteratively — Wide Research

Answers questions by researching multiple sources in parallel.

"Research passport.js JWT CVEs since January 2025"
"What changed in React Server Components v14 to v15? Focus on breaking changes"

How it works: Your query → 2-3 optimized Google searches → 30 URLs deduplicated → 12+ sources fetched → bias detection + synthesis → answer with citations.

fetch_url — Targeted Extraction

Extracts specific information from a known URL (HTML or PDF).

fetch_url(url="https://docs.stripe.com/api/authentication", looking_for="authentication methods")
fetch_url(url="https://github.com/vercel/next.js/releases/tag/v15.0.0", looking_for="breaking changes")

Quick Start

# Install
brew tap ArguSeek/arguseek && brew install arguseek
# Or: git clone ... && make install# Configureexport GOOGLE_API_KEY="your-key"export GOOGLE_CSE_ID="your-cse-id"# Run
arguseek # stdio mode (for Claude Code, Claude Desktop)

Connect to Claude Code

claude mcp add arguseek arguseek

Done. Both tools are now available to your agent.

Other Clients

Claude Desktop (~/Library/Application Support/Claude/claude_desktop_config.json):

{
"mcpServers": {
"arguseek": {
"command": "arguseek",
"env": {
"GOOGLE_API_KEY": "your-key",
"GOOGLE_CSE_ID": "your-cse-id"
}
}
}
}

HTTP mode (containers, remote):

arguseek -http # Runs on :8080
claude mcp add --transport http arguseek http://localhost:8080/mcp

Architecture

Query → Query Optimization (LLM)
→ 2-3 Parallel Google Searches
→ 30 URLs (deduplicated, ranked)
→ Fetch 12+ (two-phase fallback)
→ Bias Detection ‖ Synthesis (parallel)
→ Answer with citations

Configuration

Variable	Required	Description
GOOGLE_API_KEY	Yes	Google Custom Search API key
GOOGLE_CSE_ID	Yes	Custom Search Engine ID
GEMINI_API_KEY	No	Defaults to GOOGLE_API_KEY

Advanced

• Deployment:DEPLOYMENT.md — Cloud Run, Docker, traffic verification
• Security:PRODUCTION_SECURITY.md — Auth strategies for production
• Development:CLAUDE.md — Full API reference, architecture details

License

MIT