Learning to Use Nodes - MCP Node (MCP Tool / MCP Client)

In Agentria, you can use MCP nodes to connect externally built tools to an agent and perform real tasks — like web search or data extraction — from within a workflow. Agentria provides two MCP nodes: the MCP Tool node and the MCP Client node.

In this tutorial, you'll connect a web search MCP server so that an agent receiving a user's question can choose and call a search tool on its own, then use the result to generate an answer. At the end, you'll also see how to call a tool from the same server directly with the MCP Client node.

What Is MCP?

MCP (Model Context Protocol) is a contract that lets an agent take action. It's often compared to "a USB-C port for AI." Just as USB-C lets you connect any device with a single standard instead of a different charger for each one, MCP connects a wide range of tools to an agent through one standard specification.

An LLM can generate text, but it can't perform real actions like web searches or writing files on its own — those actions are mostly carried out through API calls. MCP is an interface that tells the agent, in a form it can read, which tool to call and with what arguments.

To understand MCP, there are three concepts to know.

• Tool — A single capability that an MCP server provides. Each tool has a description of "what it can do" and input information defining "what arguments to pass." For example, a search tool takes a query argument (the search term).

• MCP Server — A server that bundles multiple tools, receives requests from an agent, performs the real action, and returns the result. The tools a server offers vary — search, file creation, code execution, and more.

• MCP Client — The side that connects to an MCP server and calls its tools. It's a system that can call tools, combined with an LLM that decides which tool to call. The Claude desktop app is a well-known MCP client, and Agentria's MCP nodes play the same role.

The overall process works in this order.

What matters here is that the LLM is the one that decides which tool to call. Rather than fixing the call order in advance, the agent reads each tool's description and chooses whichever tool it needs at that moment. So if you connect multiple tools, the right one gets selected automatically depending on the task.

MCP isn't "documentation" — it's an "execution path." A spec document that just describes how to handle something only gives the LLM a basis for judgment; it can't produce real behavior on its own. MCP goes further than a description — it also defines the argument format to call with and the server to connect to — so when the agent sends a request in that format, the search actually runs and a result comes back.

MCP servers can run directly on your own computer (local) or be reached over the internet (remote/hosted). Agentria connects to a remote MCP server by entering its URL. So all you need is the URL to connect to and whatever credentials the server requires (such as an API key).

Before You Begin

To get started, create a project in Agentria and enter the Agent canvas.

For instructions on accessing the canvas, refer to Step 1 (Create a Project and Select a Composer) in the 🔗Core Guide: Get Started in 3 Steps guide.

After completing this tutorial, you will be able to:

1. Register MCP server information in an MCP Parameters credential.

2. Add an MCP Tool node and load the list of tools the server provides.

3. Connect the MCP Tool node to an agent node as a tool, so the agent calls tools automatically.

4. Call a specific tool directly with the MCP Client node.

Prepare an MCP Server to Connect To

To use an MCP node, you need an MCP server to connect to. This tutorial uses the Tavily web search MCP server, which is quick to sign up for and get a key from.

1. Go to https://app.tavily.com and create a free account. You get 1,000 credits a month with no credit card required.

2. Once you log in, your API key appears on the dashboard (a string starting with tvly-).

3. Fill in the issued key at below to complete the server URL. Tavily supports appending the key to the URL as a query parameter, so this single URL handles authentication too.

https://mcp.tavily.com/mcp/?tavilyApiKey=<api-key></api-key>

https://mcp.tavily.com/mcp/?tavilyApiKey=<api-key></api-key>

https://mcp.tavily.com/mcp/?tavilyApiKey=<api-key></api-key>

The Tavily MCP server provides the following tools.

If you want to use a different MCP server, look for one that offers the functionality you need on an MCP server directory and check its server URL and authentication method. Start by checking the official MCP registry for verified servers, and if you don't find what you need there, Smithery lets you look up the connection URL for hosted remote servers directly. However the server changes, connecting to it from Agentria works the same way.

Create an MCP Parameters Credential

Both MCP nodes use an MCP Parameters credential to register MCP server information. Once you create a credential, you can reuse it across multiple nodes.

The Add MCP Parameters Credential modal has the following fields.

Headers and Query Parameters are JSON input fields that default to {}. Leave them as {} if you don't need them.

Since Tavily supports appending the API key to the URL, enter the full URL you completed earlier in the MCP SERVER URL field and leave API-KEY, Headers, and Query Parameters empty.

https://mcp.tavily.com/mcp/?tavilyApiKey=<api-key></api-key>

https://mcp.tavily.com/mcp/?tavilyApiKey=<api-key></api-key>

https://mcp.tavily.com/mcp/?tavilyApiKey=<api-key></api-key>

Once you're done, click Save.

You'll get the same result if you leave the key out of the URL and instead enter {"tavilyApiKey": ""} in Query Parameters. Either way sends the same request to the server, so the single-URL approach above is simpler.

Authentication methods vary by server. For a server that uses header authentication, enter just the base address in MCP SERVER URL and put something like {"x-api-key": ""} in Headers. If the credentials are missing or wrong, the call fails at the step where you load the tool list, so enter them exactly as the server's documentation requires.

The + Add Set button at the bottom of the modal is for key rotation. If you register multiple sets, when the key currently in use hits a rate limit or runs out of tokens, calls automatically continue with the next set's key. This is useful with servers that have tight call limits, like free plans. This tutorial only uses Set 1.

Workflow Overview

The full workflow runs in the order Agent Input → Agent Loop node → Agent Output, with the MCP Tool node connected to the Agent Loop node as a tool.

The core of this tutorial is the MCP node.

The MCP Tool node loads the list of tools an MCP server provides and hands it to the agent. The agent picks whichever tool it needs from that list, fills in the required arguments on its own, and calls it. Once the search results come back, it uses them as the basis for generating a final answer.

Agentria's two MCP nodes are split as follows.

In most cases, connecting the MCP Tool node to an agent is the more convenient approach, since the agent fills in the arguments the tool call needs on its own.

MCP Tool Node Basic Info

The MCP Tool node provides the following fields.

This tutorial uses both fields.

Step 1: Check the Start Node (Agent Input)

The start node of an agent workflow is Agent Input. Double-click the Agent Input node to open the Node Editor.

Agent Input's input variables are fixed to the following four and can't be added to or edited directly.

Whatever question the user sends in chat is carried in input_message and passed to the agent as-is. So this tutorial doesn't need to declare any additional variables.

To build a workflow with input variables you declare yourself, use an Ability. An agent workflow's input is fixed to Agent Input's built-in variables.

Once you've checked this, go back to the canvas.

Step 2: Add an Agent Loop Node

Click the +Add Node button at the bottom of the canvas, and drag and drop an Agent Loop node from the AI Agent section onto the canvas.

Since the MCP Tool node doesn't run on its own and is used by connecting it to an agent node as a tool, place the agent node it will attach to first.

Step 3: Add an MCP Tool Node

From the same AI Agent section, drag and drop an MCP Tool node onto the canvas.

Placing it below the Agent Loop node makes it easier to connect as a tool later.

Step 4: Configure the MCP Tool Node (Connect the Credential & Load Tools)

Double-click the MCP Tool node to open the Node Editor.

1. In the MCP Parameters field, select the credential you created in Before You Begin.

2. Click the Connect & Load Tools button. This connects to the MCP server and loads the list of tools it provides.

3. The loaded tools (tavily-search, tavily-extract) appear in the Operation (Tool) list. Select the tools you want the agent to use. Checking Select All lets you use every loaded tool.

This step loads not just the tool names, but each tool's description and argument info as well. The agent reads this information to decide which tool to call and when. So if you select multiple tools, the agent automatically picks whichever one fits the task.

If the tool list is empty or fails to load, double-check the credential's server URL and authentication values. If the authentication values are wrong, the server rejects the request for the tool list.

Once configured, go back to the canvas.

Step 5: Connect It to the Agent Loop Node as a Tool

Connect an edge from the Tools pin at the bottom of the Agent Loop node to the In-Pin of the MCP Tool node.

Once connected as a tool, the agent automatically fills in the arguments a tool call needs (like the search query or result count), so you don't need to enter values yourself. For example, if you ask the agent to "search for what MCP is," it picks the tavily-search tool, fills query with the search term, calls it, and generates an answer from the search results that come back.

You can attach multiple MCP Tool nodes. Connecting tools from different MCP servers — search, file creation, calendar lookup, and so on — lets a single agent node handle several kinds of tasks.

Step 6: Connect the Remaining Nodes with Edges

Connect the remaining nodes with edges.

• Agent Input's Out-Pin → Agent Loop node's In-Pin

• Agent Loop node's Out-Pin → Agent Output's In-Pin

Once connected, the message the user enters in chat is passed to the Agent Loop node through input_message, and the answer the agent generates is passed on to Agent Output.

This completes the "Web Search Agent" workflow.

Step 7: Run the Whole Workflow with a Chat Test

Click the CHAT TEST button, and enter the following message in the chat input to run it.

If you get back an answer built on the search results, the workflow is working correctly. In the execution log, you can also see a record of the agent calling the tavily-search tool and the search results the server returned.

Reference: Calling a Tool Directly with the MCP Client Node

When you don't want to leave tool selection to the agent and instead want to directly specify the tool and arguments to call, use the MCP Client node. This node calls a tool on an MCP server on its own, without an agent.

Add and Configure the MCP Client Node

From the Utilities / Productivity section of the +Add Node panel, drag and drop an MCP Client node onto the canvas, then double-click it to open the Node Editor.

The MCP Client node provides the following options.

Select the Credential

In MCP Parameters, select the credential you created in Before You Begin under Select Credential. You can reuse the same credential as the MCP Tool node.

Specify the Tool to Call

Unlike the MCP Tool node, there's no agent here to fill in values for you, so you enter the tool and arguments to call directly. To run a web search on Tavily, configure it as follows.

In tool_arguments, query is the search term and max_results is the number of results to retrieve. The tool names and argument formats are defined by the MCP server you're using, not by Agentria, so check that server's documentation for the exact values to enter.

Detailed Result Flag

Turning on Detailed Result Flag (True) returns additional data, such as metadata, along with the basic result. If you only need the search result content itself, leave it off (Off).

You can also connect tool_name and tool_arguments to input variables in the Input Section. To pass in a value from a previous node, select the type to connect on the left in the Input Section, then drag and drop the variable onto each field.

Confirm It Runs with a Node Test

Click the TEST button at the top of the Node Editor to confirm the tool is called correctly.

If the call succeeds, the tool's execution result is returned in result in the Output Section.

Next Steps

🎉 Congratulations! You've successfully built the "Web Search Agent" workflow using Agentria.

Try connecting tools from several MCP servers to the MCP Tool node so the agent can handle a wider range of tasks — search, extraction, calendar management, and more — or build a workflow that uses the MCP Client node to call a specific tool with fixed arguments.

Agentria is a place where ideas become reality—your workflow can expand infinitely with your creativity.

• For more node examples, visit the 🔗Node Usage page.

Frequently Asked Questions

What is an MCP node?

An MCP node is a node in Agentria that connects to an MCP (Model Context Protocol) server to use an external tool. MCP is a contract that lets an agent take real actions like search or file creation, defining which tool to call with which arguments in a form the agent can read. Agentria provides the MCP Tool node, which connects to an agent as a tool, and the MCP Client node, which calls a tool on its own.

When should I use the MCP Tool node versus the MCP Client node?

If you want the agent to choose and call the right tool for the situation on its own, connect the MCP Tool node to an agent node like the Agent Loop node as a tool — in that case, the agent automatically fills in arguments like the search query or result count. If the tool and arguments to call are already fixed and you'll be running the same task every time, use the MCP Client node and enter tool_name and tool_arguments directly.

How is MCP different from a skill or a spec document?

A spec document describing how to handle something only gives an LLM a basis for judgment — it can't execute anything on its own. MCP goes further, defining the tool's description along with the argument format to pass and the MCP server to send the request to, so when the agent calls it according to that spec, the server actually performs the action and returns a result. In other words, if a spec document says "what needs to be done," MCP is "the path that actually executes it."

How is the MCP Tool node different from the Web Request node?

The Web Request node calls an API with a URL, method, and parameters you specify directly, so the call target and values are fixed in the workflow. The MCP Tool node first loads the tool list and each tool's description from an MCP server and hands that to the agent, which then decides which tool to call with which arguments. Use the Web Request node when the API to call is already fixed, and the MCP Tool node when the agent needs to choose a tool based on the situation.

How does the MCP Tool node work together with the Agent Loop node?

The MCP Tool node doesn't run on its own — you connect it with an edge to the Tools pin at the bottom of the Agent Loop node. Once connected, the tool list the MCP Tool node loaded is passed to the agent, and while handling the user's request, the agent selects the tool it needs, fills in the arguments, calls it, and generates an answer from the result. Connecting multiple MCP Tool nodes lets a single agent node use tools from several MCP servers together.