MCP Tools

MCP (Model Context Protocol) lets you connect your agent to external tools without writing code. Just add an mcp_url to your config and your agent gets access to tools like web search, browser automation, and more.

Quick Start

rnow init --template mcp-tavily

Get a free API key from tavily.com, then update config.yml:

rollout:
  max_turns: 3
  mcp_url: "https://mcp.tavily.com/mcp/?tavilyApiKey=YOUR_API_KEY"

Run training:

rnow run

Configuration

Single Server

rollout:
  mcp_url: "https://mcp.tavily.com/mcp/?tavilyApiKey=YOUR_KEY"

Multiple Servers

rollout:
  mcp_url:
    - "https://mcp.tavily.com/mcp/?tavilyApiKey=YOUR_TAVILY_KEY"
    - "https://mcp.exa.ai/mcp/?apiKey=YOUR_EXA_KEY"

Combined with Custom Tools

Use MCP alongside your own env.py tools:

rollout:
  mcp_url: "https://mcp.tavily.com/mcp/?tavilyApiKey=YOUR_API_KEY"

# env.py - Your custom tools
from rnow.core.tool import tool

@tool
def query_database(sql: str) -> dict:
    """Execute a SQL query."""
    return {"results": [...]}

Output: Tools: Using MCP (1 server) and env.py tools

More servers at mcp.run.

Next Steps

Train Your First Agent

Train an agent that can use tools with RL

env.py

Custom tool definitions for agents

Agent Harness

Understanding ReinforceNow's agent architecture