MCP Client Tool
MCP Client Tool
Section titled “MCP Client Tool”Connects to a remote Model Context Protocol (MCP) server and exposes the server’s tools to an AI agent. Each remote tool the model calls is gated — you approve or reject it before it runs.
What it does
Section titled “What it does”Given a remote MCP server URL, the node connects, discovers the tools the server offers, and turns each one into a tool the agent can call. When the model invokes one, the tool describes it to you (“call MCP tool search on mcp.example.com”) and runs it only after you approve. One MCP Client node can expose many remote tools at once — one approval-gated tool per remote tool.
Only remote HTTP/SSE MCP servers are supported. The extension cannot launch local (stdio) MCP servers.
When to use it
Section titled “When to use it”- You already run an MCP server (or use a hosted one) and want an agent in Agentic WorkFlow to use its tools.
- You want to reuse existing MCP tooling instead of rebuilding it as native nodes.
- You want a human approving each remote call, because MCP tools can have real side effects.
Inputs and settings
Section titled “Inputs and settings”| Setting | Notes |
|---|---|
| Server URL | The remote MCP server endpoint (HTTP/SSE), e.g. https://mcp.example.com. |
| Authentication | Optional credential sent as the Authorization header on every request. Resolved from the credential store; never logged or shown in the approval prompt. |
| Allowed Tools | Optional comma-separated allow-list of remote tool names (e.g. search, write). When set, only those tools are exposed to the agent; leave empty to expose all of the server’s tools. |
Outputs
Section titled “Outputs”Returns one or more tool dependencies (one per exposed remote tool) for agent nodes such as Tools Agent. No output port — connect it into an agent’s tools input.
Approval and safety
Section titled “Approval and safety”- Every remote call is gated. A tool runs only after you approve; the gate can’t be disabled.
- The prompt names the tool and server, never the arguments. You see the remote tool name and host — never the call arguments or the auth credential.
- Fails safe on a bad server. If the server is unreachable, the credential is invalid, or the tool list can’t be fetched, this node contributes no tools rather than breaking the agent’s other tools. Check the run logs for the connection error.
- Unattended runs auto-reject. Calls are rejected after the approval timeout when no one is present.
Dependencies and credentials
Section titled “Dependencies and credentials”- A reachable remote HTTP/SSE MCP server.
- Optional Authentication credential for servers that require an
Authorizationheader.
Example workflow
Section titled “Example workflow”Add MCP Client Tool, set Server URL to your MCP endpoint, optionally set an Allowed Tools list, and connect it into a Tools Agent. Give the agent a goal that maps to one of the server’s tools. When it calls a remote tool, an approval row appears in the Executions pane naming the tool and server; approve to let the call run.
Troubleshooting
Section titled “Troubleshooting”- No tools appear / the agent has fewer tools than expected: the server was unreachable, the credential was rejected, or the tool list failed — check the run logs. The node fails open to keep the rest of the agent working.
- Everything is rejected: a decision timed out, or you rejected the calls.
- Only some tools show: an Allowed Tools list is filtering them.
- Remember only remote HTTP/SSE servers work; local stdio servers are not supported in the extension.