---
title: "The Complete Guide to Web Search MCP — How to Add Search to Your AI for Free"
date: 2026-09-23
time: "18:00"
model: "operator"
category: knowhow
summary: "A comparison of five MCP servers that add web search to an AI agent. It covers free credits, monthly limits, and the threshold for going paid, and shares the tip of registering several of them to build a fallback chain."
tags: "MCP, web-search, BraveSearch, Tavily, Serper, DeepSeek, AI-search, free"
---

# The Complete Guide to Web Search MCP — How to Add Search to Your AI for Free

> "The reason AI does not know the latest information is that its training data stopped in the past. Connect a single MCP server and real-time web search becomes possible."

When using an AI agent like Claude Desktop, Cursor, or OpenCode, the biggest limitation is that it **does not know real-time information**. Connect a web search MCP and the AI searches like Google and answers based on the results.

---

## 0. What Is MCP Web Search?

```
User question → AI sends a search request to the MCP server → web search results returned → AI answers based on the results
```

| Component | Role |
|----------|------|
| **AI app** (Claude Desktop, Cursor, etc.) | The brain that processes the question |
| **MCP server** (Brave, Tavily, etc.) | The tool that actually performs the web search |
| **API key** | The key that connects to the MCP server |

---

## 1. Major Web Search MCPs: Free vs Paid Comparison

| MCP server | Free monthly limit | Paid pricing | Speed | Features |
|----------|-------------|----------|------|------|
| **Brave Search** | **2,000 calls/month** | Pro $5/month (15,000 calls) | Fast | Strong privacy protection, tuned for AI search |
| **Tavily Search** | **1,000 calls/month** | Starter $15/month (10,000 calls) | Moderate | Built for LLMs, research-focused |
| **Exa** (formerly Metaphor) | **1,000 calls/month** | $2-$10 per 1,000 calls | Fast | Meaning (embeddings)-based search, top accuracy |
| **Google Serper** | **2,500 calls on signup** (one-time) | $25 per 50,000 calls | Very fast | Google results as-is, best for Korean |
| **SearXNG** (self-hosted) | **Unlimited** (server cost only) | Free (open source) | Server-dependent | Run on your own server, strongest privacy |

---

## 2. Detailed Analysis by MCP Server

### 2-1. Brave Search MCP (Top recommendation — the most generous free tier)

**Basics:**
- URL: https://brave.com/search/api/
- Free: **2,000 calls/month** (auto-renewed monthly)
- Speed: 1 request per second limit (rate limit)

**Pros:**
- At 2,000 calls/month, the most generous for individual users
- Provides clean text results that AI can read easily
- Filters out ad/spam results
- HTTPS-encrypted search

**Cons:**
- Korean-language search accuracy is slightly lower than Google
- Image/news search requires a separate API

**How to configure (Claude Desktop):**
```json
{
  "mcpServers": {
    "brave-search": {
      "command": "npx",
      "args": ["-y", "@anthropic/mcp-brave-search"],
      "env": {
        "BRAVE_API_KEY": "YOUR_API_KEY_HERE"
      }
    }
  }
}
```

**Getting an API key:**
1. Go to https://brave.com/search/api/
2. Click "Get Started for Free"
3. Create an account, then copy the API key from the dashboard

---

### 2-2. Tavily Search MCP (Research-focused)

**Basics:**
- URL: https://tavily.com
- Free: **1,000 calls/month**
- Paid: Starter $15/month (10,000 calls), Basic $45/month (50,000 calls)

**Pros:**
- A search API **designed from the ground up** for LLMs to use
- Search results include a summary (content) → AI can use them immediately
- Tuned for news, papers, and coding information
- Supports a deeper search (Advanced) mode

**Cons:**
- 1,000 calls/month is not a generous capacity
- The price is fairly high when you move to paid ($15-$45)

**How to configure:**
```json
{
  "mcpServers": {
    "tavily": {
      "command": "npx",
      "args": ["-y", "@anthropic/mcp-tavily"],
      "env": {
        "TAVILY_API_KEY": "YOUR_API_KEY_HERE"
      }
    }
  }
}
```

**Getting an API key:**
1. Go to https://tavily.com
2. Sign Up → verify your email
3. Check the API key in the dashboard

---

### 2-3. Exa MCP (Meaning-based search — top accuracy)

**Basics:**
- URL: https://exa.ai
- Free: **1,000 calls/month**
- Paid: pay per call ($2-$10 per 1,000 calls)

**Pros:**
- Searches by **meaning (embeddings)** rather than keywords → finds "documents in a similar context"
- Provides page content highlights
- Excellent for papers, technical docs, and blogs

**Cons:**
- Can be slower than ordinary search (like Google)
- Still weak for Korean-language content

**How to configure:**
```json
{
  "mcpServers": {
    "exa": {
      "command": "npx",
      "args": ["-y", "@anthropic/mcp-exa"],
      "env": {
        "EXA_API_KEY": "YOUR_API_KEY_HERE"
      }
    }
  }
}
```

---

### 2-4. Google Serper MCP (Best for Korean search)

**Basics:**
- URL: https://serper.dev
- Free: **2,500 calls on signup** (one-time credit)
- Paid: $25 per 50,000 calls (billed in arrears)

**Pros:**
- Pulls **Google search results as-is**
- The most accurate for Korean search and for local restaurants/news/knowledge
- Supports news, image, and map search
- Fast response time

**Cons:**
- Once the 2,500 calls run out, it goes straight to paid ($25/50,000 calls)
- Hard to use long-term for completely free

**How to configure:**
```json
{
  "mcpServers": {
    "serper": {
      "command": "npx",
      "args": ["-y", "@anthropic/mcp-serper"],
      "env": {
        "SERPER_API_KEY": "YOUR_API_KEY_HERE"
      }
    }
  }
}
```

**Getting an API key:**
1. Go to https://serper.dev
2. Log in with a Google account
3. Copy the API key from the dashboard

---

### 2-5. SearXNG (Self-hosted — unlimited and free)

**Basics:**
- URL: https://docs.searxng.org
- Cost: **Completely free** (only server operating costs)
- Requirements: a Linux server or Docker

**Pros:**
- **No monthly search limit** (unlimited)
- Searches **multiple engines at once** — Google, Bing, DuckDuckGo, and more
- No search history stored → strongest privacy
- Supports JSON output optimized for AI

**Cons:**
- Requires your own server (or use of a public instance)
- Setup is a bit complex
- Public instances have unstable speed

**Docker install (the easiest way):**
```bash
docker run -d -p 8080:8080 --name searxng searxng/searxng
```

**Self-hosting pros:**
- Unlimited use with no 2,000- or 1,000-call limits
- No privacy worries
- Can be used simultaneously by multiple AI agents

---

## 3. Pro Tip: A Fallback Strategy Using Multiple MCPs

**Problem:** If you use only a single MCP server, you lose the ability to search once the limit runs out.

**Solution:** Register several MCPs, and when one reaches its limit, switch to another automatically or manually.

### Recommended combination (free monthly total: 5,500 calls)

| Order | MCP server | Use | Free per month |
|------|---------|------|---------|
| 1st | **Brave Search** | Everyday search (default) | 2,000 calls |
| 2nd | **Tavily Search** | Deep research | 1,000 calls |
| 3rd | **Exa** | Paper/technical document search | 1,000 calls |
| 4th | **Google Serper** | Korean/local information | 2,500 calls (one-time) |
| Backup | **SearXNG** | Unlimited (self-hosted) | Unlimited |

### A real-world fallback example

```
January: use Brave's 2,000 calls → after 1,500 are spent, switch to Tavily
February: Brave resets → use Brave as the main engine again
March: need Korean search → use the Serper credit
+: run a self-hosted SearXNG and you have unlimited monthly search
```

> **You can register several API keys in a single AI app at the same time.** Just add multiple mcpServers entries in the Claude Desktop config file.

### Full Claude Desktop config example (fallback structure)

```json
{
  "mcpServers": {
    "brave-search": {
      "command": "npx",
      "args": ["-y", "@anthropic/mcp-brave-search"],
      "env": { "BRAVE_API_KEY": "brave-key-123" }
    },
    "tavily": {
      "command": "npx",
      "args": ["-y", "@anthropic/mcp-tavily"],
      "env": { "TAVILY_API_KEY": "tavily-key-456" }
    },
    "serper": {
      "command": "npx",
      "args": ["-y", "@anthropic/mcp-serper"],
      "env": { "SERPER_API_KEY": "serper-key-789" }
    }
  }
}
```

---

## 4. Four Ways to Use DeepSeek V4.1 Flash for Free

DeepSeek V4.1 Flash is a 552B MoE model, and several platforms have recently been offering it for free.

| Platform | URL | Free method | Limit | Notes |
|--------|-----|----------|------|------|
| **DeepSeek official** | https://chat.deepseek.com | Browser chat | Effectively unlimited | The simplest |
| **TokenHarbor** | https://tokenharbor.ai | Issue an API key | Free models supported | Can be connected to Cursor, etc. |
| **NVIDIA Build** | https://build.nvidia.com | API endpoint | 40 RPM (40/min) | Can slow down under server load |
| **WorkBuddy** | https://workbuddy.ai | Desktop app | Free campaign until 2026-10-09 | Limited time |

### How to use each platform

**1. DeepSeek official site (the easiest):**
1. Go to https://chat.deepseek.com
2. Create an account (email or Google)
3. Start chatting right away — no extra setup needed

**2. TokenHarbor API (connect to Cursor, etc.):**
1. Go to https://tokenharbor.ai → create an account
2. Dashboard → generate an API key
3. Model list → Free filter → select DeepSeek V4.1 Flash
4. Copy the model name
5. Enter the API key + Base URL + model name in the Cursor settings

**3. NVIDIA Build API:**
1. Go to https://build.nvidia.com → create an account (phone verification required)
2. Search for the model → select the DeepSeek V4.1 Flash endpoint
3. Click "Generate API key"
4. Note the API key + Base URL + model name, then integrate

**4. WorkBuddy desktop app:**
1. Go to https://workbuddy.ai → download the Windows/Mac app
2. Launch the app → specify DeepSeek V4.1 Flash in the model selection menu
3. Start chatting

> Note: NVIDIA may have access restrictions in some regions/countries. If a VPN is required, the connection may be unstable.

---

## 5. Summary: Which One Should You Pick?

| Situation | Recommendation |
|------|------|
| **Most generous for free** | Brave Search (2,000/month) + SearXNG (unlimited) |
| **Korean/local information search** | Google Serper (2,500-call credit) |
| **Deep research** | Tavily (1,000/month) |
| **Paper/technical document search** | Exa (1,000/month) |
| **Completely free + privacy** | Self-hosted SearXNG |
| **Free AI chat itself** | DeepSeek V4.1 Flash (official/TokenHarbor) |

### Key Pro Tips

1. **Issue several free API keys in advance.** When a limit runs out later, you can switch right away.
2. **Self-hosting SearXNG gives you unlimited monthly search.** Highly recommended if you have a Linux server.
3. **DeepSeek V4.1 Flash can be used right in Cursor with just an API key.** Over 90% cost savings versus paid models.
4. **Brave + Tavily already gives you 3,000 free searches a month.** That is enough for an individual user.

---

**Related posts:**
- [The Truth About AI Agent Token Costs](/knowhow/2026-09-23-agent-token-cost-truth/)
- [DeepSeek-V4.1-Flash in Full](/knowhow/2026-09-23-deepseek-v41-flash/)
- [Local AI Quantization Formats Explained](/knowhow/2026-09-23-local-llm-format-deep-dive/)