MCP Tools

Extend your LLM's capabilities with built-in tools Web search, code execution, and image analysis - automatically available when using compatible models.

Overview

The Korad.AI platform supports Model Context Protocol (MCP) tools. Tools are automatically available when using compatible models, allowing your LLM to:

• Search the web for real-time information
• Execute code in a sandboxed environment
• Analyze images and extract information

Available Tools

Tool	Description	Cost
web_search	Real-time web search via Brave Search API	$0.01/query
web_fetch	Fetch and summarize URLs	$0.02/fetch
code_sandbox	Execute Python code in sandbox	$0.05/execution
image_analysis	Analyze images with vision models	$0.01/image

Tool Usage

Tools work automatically - no additional configuration needed. When you make a request that requires information the model doesn't have, it will automatically use the appropriate tool.

Example: Web Search

from openai import OpenAI

client = OpenAI(
    base_url="http://localhost:8084/v1",
    api_key="sk-bf-YOUR_VIRTUAL_KEY"
)

# The model will automatically use web_search
response = client.chat.completions.create(
    model="anthropic/claude-sonnet-4-5-20250929",
    messages=[
        {
            "role": "user",
            "content": "What's the current price of Bitcoin? Search the web."
        }
    ],
)

# The platform will:
# 1. Detect the need for web_search
# 2. Call the MCP tool
# 3. Include the tool results in the prompt
# 4. Generate the final response
# 5. Bill for tool usage + generation

Example: Code Execution

response = client.chat.completions.create(
    model="anthropic/claude-sonnet-4-5-20250929",
    messages=[
        {
            "role": "user",
            "content": "Calculate the first 100 digits of pi using Monte Carlo simulation."
        }
    ],
)

# The model will:
# 1. Recognize the need for code execution
# 2. Write Python code
# 3. Execute it in the sandbox
# 4. Return the results

Tool Configuration

Tools are configured in config.dev.json:

{
  "mcp": {
    "servers": {
      "korad-tools": {
        "command": "python",
        "args": ["./servers/korad-tools/server.py"],
        "env": {
          "BRAVE_SEARCH_KEY": "${env.BRAVE_SEARCH_KEY}",
          "BIFROST_URL": "${env.BIFROST_URL}",
          "DB_PATH": "${env.DB_PATH}"
        }
      }
    }
  }
}

Billing

Tool usage is billed separately from LLM generation:

• Tool costs appear in the X-Korad-Actual-Cost header
• Tool usage is logged in the usage database
• Tool costs are included in your monthly bill

Cost Example

X-Korad-Theoretical-Cost: $0.015000    # LLM cost
X-Korad-Tool-Cost: $0.010000            # Tool cost
X-Korad-Actual-Cost: $0.025000          # Total actual cost
X-Korad-Billed-Amount: $0.037500        # With 1.5x margin

Custom Tools

You can add custom MCP tools to your deployment:

1. Create a new MCP server in servers/your-tool/
2. Add configuration to config.dev.json
3. Restart the stack

{
  "mcp": {
    "servers": {
      "your-tool": {
        "command": "node",
        "args": ["./servers/your-tool/index.js"]
      }
    }
  }
}

Tool Availability

Tools are available for all models that support function calling:

• ✅ Claude 3.5+ (anthropic)
• ✅ GPT-4o, GPT-4 Turbo (openai)
• ✅ Gemini 2.0+ (google-gemini)
• ⚠️ DeepSeek (limited support)

Troubleshooting

Tools Not Available

If tools are not being used:

1. Check that the MCP server is running:

docker-compose logs korad-tools

2. Verify your API key for the tool (e.g., BRAVE_SEARCH_KEY)

3. Ensure your model supports function calling

Tool Errors

Tool errors are included in the response:

{
  "error": {
    "message": "Tool execution failed: web_search - rate limit exceeded",
    "type": "tool_error"
  }
}

---

Extend your LLM's capabilities with MCP tools.