🔌 Every Managed Deep Agent is an MCP server.

Let coding assistants and other agents delegate research, analysis, and more to your deployed agent as a tool.

Deploy once. Reuse across compatible MCP clients through the built-in /mcp endpoint.

Get started → https://docs.langchain.com/langsmith/python/managed-deep-agents-mcp-endpoint…

Connect MCP clients to a Managed Deep Agent - Docs by LangChain

Source: https://docs.langchain.com/langsmith/python/managed-deep-agents-mcp-endpoint Managed Deep Agents deployments run onLangSmith Agent Server, so every deployment serves the Agent ServerModel Context Protocol(MCP) endpoint at/mcp. Any MCP client that supports the Streamable HTTP transport can call the deployed agent as a tool, so another assistant or agent can delegate work to it without calling the deployment API directly.This page covers how to find the Managed Deep Agents URL and authenticate. For protocol behavior and generic Agent Server client setup, seeMCP endpoint in Agent Server.

The endpoint is the outbound direction. AnMCP connectorpoints the other way: it adds tools from a remote MCP server to your agent. A deployment can use both.

Find the endpoint URL

The endpoint is/mcpon the deployment’s API URL:

mda deployprints the LangSmith deployment dashboard URL, not the API URL. Open that dashboard, copy theAPI URLfrom the deployment details view, then append/mcp.The local server started bymda devserves the same endpoint. Append/mcpto the local server URL the CLI prints.

Authenticate requests

The MCP endpoint uses the deployment’sidentityconfiguration, the same as every other route on the deployment. Send the credential that matches the configured mode:

For the LangSmith API-key default, the key must belong to the workspace that owns the deployment. A key that is valid for another workspace is rejected.Authentication failures return one of three responses:

The local server started bymda devrequires no credential. It grants every caller an unscoped local service principal, so any client on the machine can call/mcpand every other route. Only a deployment enforces the header.

Understand what the agent exposes

The Agent Server exposes the agent as an MCP tool:

• Tool name: The agentnameset in theagent definition.
• Tool input schema: The agent’s input schema.

For transport details and other Agent Server MCP behavior, seeMCP endpoint in Agent Server.

Add the agent to Claude Code

Register the endpoint as an HTTP MCP server:

Claude Code lists the agent as a tool once the server connects.In a JSON configuration, Claude Code expands$\{VAR\}in header values, so the key stays in the environment instead of the config file:

Add the agent to another MCP client

Clients that read anmcpServersconfiguration accept the endpoint and header directly:

A client that accepts only a URL and negotiates OAuth cannot authenticate to the endpoint, which requires a header credential. Configure the header explicitly. Without one, a client may fall back to OAuth and report a Dynamic Client Registration failure rather than the missing header, which obscures the real cause.

Call the endpoint from code

Load the agent’s tool through an MCP client library, then pass the tools to a model or another agent:

Read credentials from the environment. Do not hard-code them in client code.

Understand session behavior

Agent Server MCP requests arestateless: separate calls do not share thread state.To retain knowledge across calls,opt in to durable memory. Durable memory is optional and shared by every caller of the deployment. Do not enable it when MCP clients should not influence one another.

When to use the MCP endpoint

Next steps