MCP Tools

Config agents support MCP (Model Context Protocol) tool integrations. MCPs are external tool servers that expose tools the LLM can call during task execution. Society AI manages the MCP server lifecycle and API keys -- you just toggle them on.

How MCP Integration Works

When you enable an MCP for your agent:

1. The MCP server is started as a subprocess using the configured command (typically npx).
2. The server's tools are automatically registered with the LLM via PydanticAI's toolset system.
3. During task execution, the LLM can call any of the MCP's tools as needed.
4. Tool call costs are tracked and added to the infrastructure cost for billing.

MCP tools appear in the agent's system prompt under the Capabilities section, so the LLM knows what tools are available and when to use them.

Enabling MCPs

Add MCP identifiers to the enabled_mcps list in your agent's configuration:

{
  "enabled_mcps": ["brave_search", "firecrawl"]
}

In the Agent Builder UI, toggle MCPs on or off. The changes take effect on the next task execution.

Available MCPs

Web Search (brave_search)

Search the web using the Brave Search API. Returns current information, news, and research results.

The brave_web_search tool accepts a search query and returns web results with snippets and URLs. Use this when your agent needs access to current information that is not in its training data.

{
  "enabled_mcps": ["brave_search"]
}

Web Scraper (firecrawl)

Scrape, crawl, and extract structured data from websites. Supports single-page scraping, multi-page crawling, URL discovery, and structured data extraction.

Firecrawl provides multiple tools, each with its own pricing:

{
  "enabled_mcps": ["firecrawl"]
}

CoinGecko Crypto Data (coingecko)

Access real-time and historical cryptocurrency data from CoinGecko. Get prices, market caps, trading volumes, trending coins, exchange data, NFT collections, and on-chain DEX data.

CoinGecko provides three tools:

The invoke_api_endpoint tool is a universal gateway to the CoinGecko API. The LLM first discovers available endpoints, then calls the appropriate one with the right parameters.

{
  "enabled_mcps": ["coingecko"]
}

Combining MCPs

You can enable multiple MCPs simultaneously. The LLM will have access to all tools and can use them in combination:

{
  "enabled_mcps": ["brave_search", "firecrawl", "coingecko"]
}

When multiple MCPs are enabled, the system prompt lists all available tools under the Capabilities section, helping the LLM choose the right tool for each query.

Cost Tracking

MCP tool calls are tracked and included in the infrastructure_cost metadata returned with every response:

{
  "infrastructure_cost": {
    "llm_cost_usd": "0.0023",
    "mcp_tools": {
      "brave_web_search": {
        "mcp_id": "brave_search",
        "count": 2,
        "cost_per_call": 0.005,
        "total_cost": 0.01
      }
    },
    "mcp_total_usd": "0.01",
    "total_usd": "0.0123"
  }
}

Each tool call is recorded with its MCP ID, call count, per-call cost, and total cost. This allows you to understand exactly what your agent is spending on external tool usage.

Writing Instructions for MCP Tools

When you enable MCPs, add instructions that guide the LLM on when and how to use them:

# With brave_search enabled
Use web search for any question about current events, prices, or news.
Do not search for information you already know from training data.

# With firecrawl enabled
When the user provides a URL, scrape it to get the full content.
For research tasks, search for relevant URLs first, then scrape the best results.

# With coingecko enabled
Use CoinGecko for any cryptocurrency price or market data questions.
Always include the data timestamp in your response.

These instructions are added to the system prompt and help the LLM make better decisions about tool usage.