Agent tools and MCP

Orchestration settings

Maximum agent steps

The maximum number of steps the agent is allowed to take before returning a response. Each step is a separate request to the LLM provider and may involve reasoning or calling a tool. Set a lower value to reduce compute cost and prevent run-away loops; set it higher to allow the agent to complete more complex, multi-step tasks.

Memory tool

Enabling this appends a Memory tool to the agent.

When enabled, the agent gains access to a Memory tool that lets it search through the user’s previous conversations with this agent. Use this to give the agent awareness of past interactions — it can recall context, facts, or information from earlier threads, and always references the conversation title and date when presenting results.

When Agent Memory is enabled in the system settings and a Global Embedding model is configured, the memory tool searches past conversations using vector similarity. Otherwise, it falls back to a keyword-based search.

Dynamic tool selection

Enabling this appends a Tool Selection tool to the agent.

When enabled, the first request to the LLM includes only a single tool: a tool that lists all available tools and their descriptions. The agent uses this to declare which tools it actually needs for the task. In all subsequent steps, only the selected tools are included in the request to the LLM provider.

This reduces the number of tool definitions sent on each request, which lowers token usage and can improve decision accuracy by avoiding unnecessary tool calls.

Core tools

This equips the agent with external tools that extend its functionality, such as reading data from a database, sending emails, generating PDFs, or executing scripts. When tools are available, the agent can decide which ones to use, call them in sequence to solve complex tasks, and retry or choose alternative tools if an operation fails.

Use the edit button in the top-right corner of the section to manage the tools assigned to this agent.

Learn more about the available tools and how to configure them in AI Tools.

Whether tools are called correctly and with fitting parameters may depend on the LLM used. Modern LLMs are fine-tuned to be good at calling tools.

Model Context Protocol (MCP)

Naia Agent Studio supports connecting tools to agents via the Model Context Protocol (MCP).

This enables your AI agents to discover and use capabilities across various services through a standardized interface. You can connect locally (npm package) or to a remote server.

When connecting to a server or local file, the resource responds with a list of tools which are then equipped to the agent.

Local MCP connection

Open the Node Package Manager.
Add a package that has MCP support.
Provide an MCP Executable Path. This is the relative path within the project that points to the main file that serves as the entry point for the MCP connection.
Save.
Select the connection button to test the connection.

Remote MCP connection

Open the Remote Systems tool.
Add a new system and select type Model Context Protocol.
Select the protocol. This depends on the server.
Enter a URL.
Optional: Configure authentication for the remote system.
Import the tools.

By importing the tools, the metadata of the remote server’s tools are saved in the database to avoid connecting to the remote server on every request. Re-import whenever you need to pick up the latest changes to the tool definitions.

If the server requires a header value or an authorization key, you can provide authentication to the remote system by creating a proxy authentication.

Add an MCP connection to your agent

Select your agent.
In the Tools & MCP tab, use the edit button next to Local MCP Connections or Remote MCP Connections to add a connection.
1. In each remote connection, use the Tool Selection button to choose which of the imported tools the agent should use.
Save.

Result: Your agent can access the MCP tools and use them to perform tasks based on the capabilities provided by the connected servers.