MCP Call Node — Node Reference

The MCP Call node calls a specific Model Context Protocol (MCP) tool directly, without an LLM deciding which tool to invoke. Use it when you know exactly which tool to call and want deterministic, single-step execution.

Overview

Property	Value
Inputs	1
Outputs	1
Output	`$nodeLabel.result`

When to Use

Use the MCP Call node when:

You know exactly which MCP tool to call at design time
You do not need an LLM to choose between tools
You want a lightweight, deterministic alternative to an agent node for a single tool invocation

Use the Agent node instead when an LLM should decide which tools to call or when you need multi-step reasoning.

The MCP Call node cannot be connected to an agent node as a tool.

Connection Configuration

The node connects to a single MCP server. Choose a transport and fill in the corresponding fields.

SSE / Streamable HTTP

Field	Description
`transport`	`sse` or `streamable_http`
`url`	Full URL of the MCP server (e.g. `https://mcp.example.com/sse`). Supports DSL expressions.
`headers`	JSON object of HTTP headers (e.g. `{"Authorization": "Bearer ..."}`)
`timeoutSeconds`	Per-connection timeout in seconds (default: 30)

stdio

Field	Description
`transport`	`stdio`
`command`	Executable to run (e.g. `npx`)
`args`	JSON array of arguments (e.g. `["-y", "@modelcontextprotocol/server-filesystem"]`)
`env`	JSON object of environment variables (e.g. `{"API_KEY": "..."}`)
`timeoutSeconds`	Per-connection timeout in seconds (default: 30)

Tool Selection

Fill in the connection fields.
Click Fetch Tools — the node connects to the MCP server and lists available tools.
Select a tool from the dropdown.

Tool selection is required. The node will not execute without a selected tool.

Arguments

After selecting a tool, the node renders one input field per argument defined in the tool's input schema. Each field accepts:

A static value: 10
A DSL expression: $userInput.body.text

Required arguments are marked with *.

Output

Field	Description
`$nodeLabel.result`	Tool result. If the response is valid JSON, this is a parsed object. Otherwise it is a string.

Example

Call a search tool on an SSE MCP server, passing the user's query:

{
  "type": "mcpCall",
  "data": {
    "label": "searchCall",
    "connection": {
      "transport": "sse",
      "url": "https://mcp.example.com/sse",
      "timeoutSeconds": 30
    },
    "selectedTool": "search",
    "toolArguments": {
      "query": "$userInput.body.text",
      "limit": "10"
    },
    "timeoutSeconds": 30
  }
}

Access the result downstream:

$searchCall.result.items   ← if result is a JSON object
$searchCall.result         ← if result is a plain string

Notes

The node resolves DSL expressions in url, headers, env, and all toolArguments values before calling the tool.
The timeoutSeconds at the node level applies to the overall execution; the connection.timeoutSeconds applies to the MCP session itself.
If the MCP server is unavailable or returns an error, the node fails and the error message is available in the workflow trace.