MCP Tools

This example connects a remote MCP tool server, discovers tools from it, and runs those tools inside the normal smart runtime.

Open source: examples/mcp-tavily/mcp_tavily.ts

Use this when

• tool capabilities live outside your process
• you want remote tools to participate in the same planning and tracing loop as local tools
• you need a real MCP integration reference instead of a conceptual guide

What it shows

• MultiServerMCPClient setup
• remote tool discovery
• fromLangchainTools(...) bridge into Agent SDK
• smart runtime planning over MCP tools

Run it

cd examples
npm run example:mcp-tavily

Core code

const client = new MultiServerMCPClient({
	throwOnLoadError: true,
	prefixToolNameWithServerName: true,
	useStandardContentBlocks: true,
	mcpServers: {
		"tavily-remote-mcp": {
			transport: "stdio",
			command: "npx",
			args: ["-y", "mcp-remote", `https://mcp.tavily.com/mcp/?tavilyApiKey=${TAVILY_API_KEY}`],
			env: {},
		},
	},
});

const tools = fromLangchainTools(await client.getTools());
const agent = createSmartAgent({
	model,
	tools,
	useTodoList: true,
	limits: { maxToolCalls: 10 },
	summarization: { enable: true, maxTokens: 6000 },
});

Environment notes

This example expects the relevant provider credentials, including the Tavily API key and the model provider key used by the script.

End-to-end flow

1. An MCP client is created with a Tavily-backed server definition.
2. The client discovers remote tools.
3. fromLangchainTools(...) adapts them into SDK-native tools.
4. A smart agent uses those tools with planning and summarization enabled.
5. The client is closed in a finally block after the run.

Why this example matters

It proves that MCP tools do not need a separate orchestration path. They can live inside the same planning, tracing, and summarization loop as local tools.

What to inspect

• the MCP client bootstrap
• tool discovery through client.getTools()
• conversion with fromLangchainTools(...)
• cleanup with client.close() in the finally block

Production takeaway

This is the reference pattern when MCP is part of the production tool surface rather than a separate experimental path.

Expected output

• discovered MCP tool names are printed before the run
• if credentials are present, the agent returns a short Tavily-backed summary
• if keys are missing, the script explains why the agent run is skipped

Common failure modes

• OPENAI_API_KEY or TAVILY_API_KEY is missing
• the remote MCP transport fails to connect or load tools
• the example is interrupted before client.close() executes, leaving cleanup incomplete