Tools and MCP (LangGraph sample)

MCP behavior when enabled

DataRobot can attach extra tools (including from an MCP deployment) when the agent runs. In the graph sample, those appear alongside tools defined in Python. If only local tools are listed and the injected list is ignored, MCP tools will not be available to the model.

Practical rule: merge platform tools with local tools wherever tools bind to the graph (the e2e myagent.py shows the intended pattern).

Configuration outside the repo

MCP server URL and credentials come from the deployment / environment, not from the LangGraph workflow.yaml in the minimal sample. The NAT workflow example adds function_groups in YAML instead; see nat/mcp.md. With MCP_WORKLOAD_ID, the agent asks the platform where that workload is served (GET /api/v2/workloads/<id>/) and appends /mcp to the endpoint it reports, so the same variable works whether the cluster routes workloads through different API Gateway. The agent's API token therefore needs read access to the workload. If the lookup cannot answer, the agent runs without MCP tools and logs a warning. Then, no URL is guessed, because a composed one would be wrong on some clusters. Use EXTERNAL_MCP_URL to address a workload directly instead.

Automated tests

e2e-tests/dragent_tests/test_mcp.py exercises MCP tool calls when either MCP_DEPLOYMENT_ID or MCP_WORKLOAD_ID is set and a tool-capable agent is configured.