Skip to main content

​
Overview

The MCP Node lets an Agent use tools exposed by a remote MCP server. It acts as a tool provider, not a normal execution step in the main flow.

​
What Users Can Do

With MCP connected to an Agent, users can:
  • call external tools exposed by an MCP server
  • augment Agent responses with real-time/tool-backed data
  • combine MCP tools with internal tools (query, mutation, http, code, etc.)
Common use cases:
  • web search (for example Exa MCP)
  • retrieval from external systems
  • domain-specific tool catalogs exposed via MCP

​
Required Configuration

  • Label - display name
  • MCP URL - server endpoint (required)
  • Auth Type - none | bearer
  • Bearer Token - optional/required depending on server auth

​
Connection Rules (Canonical)

MCP must be connected through Agent tools using this direction: 
{
  "source": "agentNodeId",
  "sourceHandle": "tools",
  "target": "mcpNodeId",
  "targetHandle": "input"
}
Rules:
  • MCP should only be connected through Agent tools edges.
  • MCP is not part of the primary execution dependency chain.
  • Tools edges are separate from normal input/output orchestration edges.

​
Runtime Behavior

When Agent runs:
  1. MCP client is created using MCP URL/auth
  2. tools are discovered from mcpClient.tools()
  3. discovered MCP tools are merged into Agent tool map
  4. MCP client is closed during finalization/cleanup

​
Example: Exa MCP

  • URL: https://mcp.exa.ai/mcp
  • Auth Type: none (or bearer if your server requires token auth)
  • Connect:
    • agent.tools -> mcp.input

​
Example: MCP with Bearer Auth

MCP node: 
{
  "nodeType": "mcp",
  "label": "Private MCP",
  "url": "https://your-private-mcp.com/mcp",
  "authType": "bearer",
  "bearerToken": "YOUR_MCP_BEARER_TOKEN"
}
Tools edge: 
{
  "source": "agentNodeId",
  "sourceHandle": "tools",
  "target": "mcpNodeId",
  "targetHandle": "input"
}
Notes:
  • Keep bearer token in environment/secret management; avoid hardcoding in shared templates.
  • If your MCP token rotates, update node config and redeploy/re-generate as needed.

​
Tips

  • Keep Agent toolConfig.maxSteps sufficient for multi-step tool calls (default 10).
  • If old flows were created before canonical tools direction, reconnect MCP edges to match:
    • sourceHandle: "tools" on Agent
    • targetHandle: "input" on MCP