Connect MCP tools to Managed Deep Agents

Managed Deep Agents use MCP connectors to load tools from remote MCP servers. Declare the servers in connectors/mcp.ts or connectors/mcp.py, export a named mcp declaration, and Managed Deep Agents loads those tools into the agent at runtime.

Managed Deep Agents is in private beta, available on LangSmith Cloud in the US region only. Join the waitlist to request access.

For other connector types, see Connectors.

Add an MCP connector

Add connectors/mcp.py or connectors/mcp.ts next to your agent entry file. The connector module must export a named mcp declaration.

from managed_deepagents.connectors import define_mcp_servers

mcp = define_mcp_servers(
    mcp_servers={
        "langchainDocs": {
            "transport": "http",
            "url": "https://docs.langchain.com/mcp",
        },
    },
)

import { defineMcpServers } from "managed-deepagents";

export const mcp = defineMcpServers({
  mcpServers: {
    langchainDocs: {
      transport: "http",
      url: "https://docs.langchain.com/mcp",
    },
  },
});

You do not import MultiServerMCPClient or call getTools() / get_tools() yourself. mda discovers the connector module, injects the MCP adapter dependency into the compiled build, creates the client in the managed runtime, loads the tools, and appends them to the authored tools from agent.ts or agent.py.

Supported MCP servers

Connectors support remote MCP servers only:

Transport	Use
`http`	Streamable HTTP MCP servers.
`sse`	Legacy SSE MCP servers.

Stdio MCP servers are not supported in connectors. If a server needs local process management, expose it over HTTP/SSE or wrap the behavior as a normal authored tool.

Configure server options

Each server key is the logical server name Managed Deep Agents uses for validation, tracing metadata, and tool-name prefixing. Server configs can include static headers. Connectors do not run an OAuth authorization flow. If an MCP server requires OAuth, provide a pre-provisioned access token or another static credential through headers. Store the token in .env so mda dev can load it locally and mda deploy can forward it as a hosted deployment secret. The connector module is normal project code, so read secrets as environment variables with os.environ in Python or process.env in TypeScript. You do not import the .env file directly.

import os

from managed_deepagents.connectors import define_mcp_servers

mcp = define_mcp_servers(
    mcp_servers={
        "github": {
            "transport": "http",
            "url": "https://example.com/mcp",
            "headers": {
                "Authorization": f"Bearer {os.environ['GITHUB_MCP_TOKEN']}",
            },
        },
    },
)

import { defineMcpServers } from "managed-deepagents";

export const mcp = defineMcpServers({
  mcpServers: {
    github: {
      transport: "http",
      url: "https://example.com/mcp",
      headers: {
        Authorization: `Bearer ${process.env.GITHUB_MCP_TOKEN}`,
      },
    },
  },
});

Do not commit MCP tokens, API keys, OAuth access tokens, or passwords. Put local values in .env; mda dev loads them for local development, and mda deploy forwards non-reserved .env values as hosted deployment secrets.

MCP connector defaults

Managed Deep Agents applies managed defaults when it loads connector tools:

Option	Default	Description
`prefixToolNameWithServerName` / `prefix_tool_name_with_server_name`	`true`	Prefix MCP tool names with the server name, for example `github__search`, to avoid collisions.
`throwOnLoadError` / `throw_on_load_error`	`true`	Fail when tools cannot be loaded instead of starting with a partial tool surface.
`useStandardContentBlocks` / `use_standard_content_blocks`	`true`	Convert MCP tool outputs to standard LangChain content blocks. Python connectors currently require the default `true` value.
`onConnectionError` / `on_connection_error`	`"throw"`	Fail when a server cannot be reached. `"throw"` is the only supported value.

Disable tool-name prefixing only when you know the MCP tool names do not collide. Managed Deep Agents validates duplicate MCP tool names when prefixing is disabled.

Test and deploy

Test the project locally with mda dev, then deploy it with mda deploy. Open deployment traces in LangSmith to inspect model calls, tool calls, errors, and latency. MCP misconfiguration surfaces during local startup or first tool load, depending on when the runtime reaches the MCP server.

Next steps

Connectors

Compare MCP and LangSmith connector types.

Custom tools

Add authored tools alongside MCP connector tools.

Deploy an agent

Run and deploy the connector-enabled agent.

Connect these docs to Claude, VSCode, and more via MCP for real-time answers.

Edit this page on GitHub or file an issue.

​Add an MCP connector

​Supported MCP servers

​Configure server options

​MCP connector defaults

​Test and deploy

​Next steps

Connectors

Custom tools

Deploy an agent

Add an MCP connector

Supported MCP servers

Configure server options

MCP connector defaults

Test and deploy

Next steps