MCP

Model Context Protocol is a strong fit for autonomous agents because it turns external capabilities into discoverable tools instead of hard-coded SDK integrations.

What MCP gives you here

With Agent SDK, MCP is mainly about one thing: turning remote tool servers into normal runtime tools that participate in the same planning, tracing, approvals, and summarization flow as local tools.

Recommended setup

import { createSmartAgent, fromLangchainModel, fromLangchainTools } from "@cognipeer/agent-sdk";
import { ChatOpenAI } from "@langchain/openai";
import { MultiServerMCPClient } from "@langchain/mcp-adapters";

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=${process.env.TAVILY_API_KEY}`],
			env: {},
		},
	},
});

const tools = fromLangchainTools(await client.getTools());
const model = fromLangchainModel(new ChatOpenAI({ model: "gpt-4o-mini", apiKey: process.env.OPENAI_API_KEY }));

const agent = createSmartAgent({
	name: "MCP Explorer",
	model,
	tools,
	runtimeProfile: "balanced",
	planning: { mode: "todo" },
	limits: { maxToolCalls: 10, maxContextTokens: 12000 },
});

Why this matters for autonomous agents

MCP becomes especially valuable when the agent needs to:

• discover tools from a remote server at runtime
• mix local and remote tools in one execution loop
• keep external actions inside the same approval and trace pipeline
• recover after large tool outputs are summarized

Best practices

• prefer planning.mode over legacy useTodoList
• prefix server tool names so multi-server environments stay unambiguous
• expect some MCP tools to return large payloads and plan for summarization
• use tracing from day one so remote-tool failures are visible

Common failure points

Tool names are unclear

Enable server-prefixed names when you connect to multiple MCP servers.

Outputs are too large

Use smart runtime summarization and get_tool_response for recovery instead of forcing huge raw payloads into every turn.

Authentication is flaky

Treat MCP credentials like any other production secret and keep the auth surface outside prompts.

Example to run

The repository includes an MCP example:

cd examples
npm run example:mcp-tavily