LangSmith/Deployment/Q/How do deployed agents work together over the MCP endpoint?

• How do deployed agents work together over the MCP endpoint, including from Claude/Code or Codex?

  • AI Answer

    • Answer/Official from Agent Server: MCP endpoint, researched by Anthropic/Model/Claude/5/5/Opus on 2026-09-29 Tue
    • Short answer: every LangSmith/Deployment serves /mcp, and each agent in it shows up there as one MCP tool. Any MCP client that speaks Streamable HTTP — another agent, Claude/Code, Codex — can list those tools and call an agent like any other tool.

    • What the endpoint exposes

      • The Agent Server implements MCP over the Streamable HTTP transport at /mcp. It is on by default in langgraph-api>=0.2.3; "disable_mcp": true in langgraph.json turns it off.
      • Each deployed agent becomes one tool: the tool name is the agent's name, the description is the agent's description, and the input schema is the agent's input schema. Name and description are set in langgraph.json.
      • The docs recommend explicit, minimal input and output schemas. The default MessagesState accepts any message type, which is too loose for another model to fill in well.
      • Auth is the same as the rest of the deployment's API; on LangSmith that is an X-Api-Key header.
      • The endpoint has no sessions. Every /mcp call is independent, so one tool call is one fresh run of the agent, with no thread carried to the next call.

    • Agent to agent

      • An agent in one deployment can load another deployment's /mcp tools with langchain-mcp-adapters (load_mcp_tools or MultiServerMCPClient) and hand them to its own model. The calling agent's model then decides when to delegate to the other agent, the same way it picks any tool.
      • That makes deployed agents interchangeable: swapping one for another means pointing the client at a different URL.
      • Because calls are stateless, MCP suits one-shot delegation: ask a question, get an answer.
        • For long-running or multi-turn work between agents, the docs point to other routes: A2A, served per assistant at /a2a/{assistant_id} with tasks and streaming (A2A endpoint), or LangSmith/Deep Agents async subagents, which launch, check, steer and cancel runs on a remote deployment over the Agent Protocol (Async subagents).

    • From Claude/Code

      • bash

        claude mcp add --transport http finance-agent \
          https://<deployment-host>/mcp \
          --header "X-Api-Key: $LANGSMITH_API_KEY"

      • The agent then appears to Claude Code as a tool named after the agent.

    • From Codex

      • codex mcp add --url registers a Streamable HTTP server, but its only auth flag is --bearer-token-env-var, which sends Authorization: Bearer. LangSmith expects X-Api-Key, so set the header in ~/.codex/config.toml:
      • toml

        [mcp_servers.finance-agent]
        url = "https://<deployment-host>/mcp"
        env_http_headers = { "X-Api-Key" = "LANGSMITH_API_KEY" }

      • env_http_headers maps a header name to the environment variable that holds its value. Checked against the codex mcp add --help output and the config keys of codex-cli 0.158.0, not against a live deployment.

    • User-scoped tools

      • Going the other way, a deployed agent can call MCP servers on behalf of each user. Custom auth middleware fills langgraph_auth_user, and a node builds a MultiServerMCPClient with that user's token in the request headers, so each user gets only their own tools.