Skip to content
threadle

MCP tool node

◈ MCP tool is the canvas client for MCP. Each run connects to the chosen server, calls one tool, and tears down (same timeout / output-cap spirit as custom nodes). Stdio servers are spawned as children. HTTP/SSE use the configured URL.

For exposing your workflows to hosts, see MCP (threadle mcp).

Add one

Section titled “Add one”
  1. Palette Nodes → MCP tool, or pick it from the wire-drop menu.
  2. Select the node → inspector:
  • server. Discovered from project / threadle registry / host configs (disabled servers are hidden)
  • tool. Listed after connecting once
  • arg port. Optional. Inbound text fills this argument
  • params. Schema-driven fields for the rest of the args
  1. Wire text in (optional) and text out → ▶ Run.

Glyph ◈, wire colour --wire-mcp.

If the list is empty: enable MCP client in Settings, refresh servers, or add a config (see discovery).

Data shape

Section titled “Data shape”
{
  type: "mcp-tool";
  ref: { server: string };   // discovery id
  tool: string;
  params?: Record<string, string>;
  argPort?: string;
  snapshot?: { params?, toolDescription?, serverLabel? };
}

When to use it

Section titled “When to use it”
Use MCP tool when…Prefer instead when…
You want a deterministic, token-free step (ticket fetch, DB query, Slack post via an existing MCP server)The work is “ask an agent” → use an agent node
The integration already exists as an MCP serverYou own a small script → custom node
You are composing host tools into a larger graphYou need the host to call your graph → MCP server
Section titled “Related”