build your own research tool in less than 20 minutes

Build a simple AI research tool that fits your needs.

Jun 07, 2025

After trying out Claude and Google Gemini’s research features, I found that they are not very research-friendly.

The problem? These research features are black boxes. No control. No guidance. Just watch helplessly as the AI chases every tangent it finds.

The solution? Build your own research assistant that YOU control. One that searches and scrapes in a single API call. One that actually stays on topic.

After my rant about full-auto AI research tools last week (here if you missed it), I'm showing you the alternative I built.

Let’s create our own little research assistant.

What we're building:

✅ Human-in-the-loop research (you guide, AI executes)

✅ Search + scrape in ONE API call

✅ No more "AI went rogue" moments

✅ Actually delivers what you asked for

Download the complete n8n template: Search-Scrape-Firecrawl-AI-Agent.json

Download the complementary sub-workflow: Firecrawl-Scrape-Tool.json

Firecrawl's New Search Endpoint

While everyone's still duct-taping search APIs to scraping tools, Firecrawl quietly released an endpoint that handles both in one request.

Firecrawl approach:

Let's Build This Thing (20 Minutes, I Timed It)

What You Actually Need

The non-negotiables:

n8n instance (cloud or self-hosted - I use Render)
Firecrawl API key (free tier available)
OpenRouter account (for AI model access)
Basic n8n knowledge (if you can drag and drop, you're golden)

Real talk: If you're still on the fence about self-hosting n8n, this one workflow pays for your server 20x over. Stop overthinking it.

Step 1: Create Your AI Research Interface

First up, we need a way to chat with our research assistant. n8n's chat trigger creates a clean web interface faster than you can say "another SaaS subscription."

The setup:

Drag in a "When chat message received" trigger
Enable "Public" (creates a shareable URL)
That's it. No, really.

Pro tip from the trenches: I embed this chat directly into my Notion workspace using the webhook URL. Research without tab-switching. Game changer.

Step 2: Configure the AI Brain (Where Shit Gets Real)

The AI Agent node is the conductor of this beautiful chaos. It takes your question, figures out what to search for, and synthesizes the results into something actually useful.

The system prompt that makes it work:

You are a helpful assistant that answers questions based on information found on the web. 

You will use the firecrawl_tool to answer the question the user has. You only need to use it once. 

Grab the results, analyze them and answer the questions. Provide the 3 URLs analyzed for fact checking.

The tool that you have access to requires a query. Formulate one first and feed it into the tool, provide the output as JSON:

{ Query: {{ querycontents }} }

Why this exact prompt matters:

"You only need to use it once" → Stops expensive search loops
"Provide the 3 URLs" → No more hallucinations
JSON output requirement → Keeps data flow clean

Step 3: Pick A Model thats Good with Tool Use

I'm using Google's Gemini 2.5 Pro through OpenRouter. Here's why:

Gemini 2.5 Pro supports tool use, meaning it can call external functions, generate structured output (like JSON), execute code, and use search. This allows the model to solve multi-step tasks, call APIs, or format responses for specific downstream systems.
Source: Gemini API Docs

Tool use is important because the agent needs to use the Firecrawl search tool to get the context needed.

Step 4: Build the Firecrawl Search Tool

This sub-workflow is where your assistant actually searches and scrapes. We separate it so the AI can call it intelligently instead of randomly.

Main workflow setup:

Description: "Call this tool to search the web and answer queries"
Workflow: Link to your Firecrawl Search sub-workflow
Input: Map the chat query to the search query

Step 4a: The Sub-Workflow That Does the Actual Work

Workflow trigger config:

Input parameter: "query"
This catches the search query from your main workflow

Step 4b: The Firecrawl API Configuration (Copy This Exactly)

Here's where we tap into that beautiful search endpoint:

Method: POST
URL: https://api.firecrawl.dev/v1/search
Auth: Bearer Token (your Firecrawl API key)

Body:
{
  "query": "{{ $json.query }}",
  "limit": 3,
  "lang": "en",
  "country": "us",
  "scrapeOptions": {
    "formats": ["markdown", "links"]
  }
}

Why these exact settings:

Limit 3: Perfect balance of comprehensive and affordable
Markdown format: AI models eat this up 10x better than HTML
Country/lang: Keeps results relevant (adjust for your market)

Step 5: Test This Beautiful Monster

Time to watch your money-saving robot come alive:

Testing sequence:

Grab your chat URL (from the webhook)
Ask: "What are the best n8n alternatives in 2024?"
Watch your workflow execute in real-time
Try not to giggle at how fast it works

What winning looks like:

Response time: 15-25 seconds
Answer includes: Solid summary + 3 source URLs
Information: Current and actually accurate
Cost per query: ~$0.03-0.06

When Things Go Sideways (And How to Fix Them)

Empty or generic responses

The tool connection is borked
Check: Workflow ID matches your sub-workflow exactly

Timeout on complex queries

Some sites are bloated messes that take forever to scrape
Fix: Add timeout settings or drop limit to 2

Costs going through the roof

Your AI is search-happy and hitting the API repeatedly
Verify: "use it once" is in your system prompt

Advanced Tricks That Make Me Look Smart

Turbocharge Your Research Quality

{
  "query": "your search",
  "limit": 5,  // Go deeper when needed
  "scrapeOptions": {
    "formats": ["markdown", "links", "html"],
    "includeImages": true,  // For visual research
    "maxDepth": 2  // Scrape linked pages too
  }
}

Give Your Assistant a Memory

I hooked mine up to Supabase to:

Store all queries and responses
Build a searchable knowledge base
Track research patterns
Skip duplicate searches (save those pennies)

Integration Ideas That Actually Make Money

Slack Research Bot:

Team asks questions in Slack
Bot drops researched answers
Saves 10+ hours/week of "let me Google that"

Content Research Pipeline:

Feed it blog topics
Get comprehensive research reports
Include competitor analysis
Everything pre-formatted

Sales Intelligence Machine:

Research prospects automatically
Pull recent company news
Find decision-maker insights
Your sales team looks psychic

Your Move

This automation changed my entire workflow. No more 47 browser tabs. No more "I'll research that and circle back."

The template's ready. Setup takes 20 minutes (I literally timed it). The savings hit immediately.

With this, you're building a research system that scales infinitely without your costs scaling with it.

Download the template. Set it up tonight. Thank me when you see the results!

🔓 Get All Future Automation Blueprints Free. Never miss a money-saving workflow.