NVIDIA NeMo Agent Toolkit as an MCP Client#

Model Context Protocol (MCP) is an open protocol developed by Anthropic that standardizes how applications provide context to LLMs. You can read more about MCP here.

You can create a workflow that uses MCP tools as functions. In this case, the workflow acts as an MCP host and creates MCP clients to connect to MCP servers and use their tools as functions.

This guide covers how to use a NeMo Agent toolkit workflow as an MCP host with one or more MCP clients. For more information on how to use the NeMo Agent toolkit as an MCP server, refer to MCP Server.

Installation#

MCP client functionality requires the nvidia-nat-mcp package. Install it with:

uv pip install "nvidia-nat[mcp]"

Accessing Protected MCP Servers#

NeMo Agent toolkit can access protected MCP servers through the MCP client auth provider. For more information, refer to MCP Authentication.

MCP Client Configuration#

NeMo Agent toolkit enables workflows to use MCP tools as functions. The library handles the MCP server connection, tool discovery, and function registration. This allows the workflow to use MCP tools as regular functions.

Tools served by remote MCP servers can be used as NeMo Agent toolkit functions using mcp_client, a flexible configuration using function groups that allows you to connect to an MCP server, dynamically discover the tools it serves, and register them as NeMo Agent toolkit functions.

`mcp_client` Configuration#

function_groups:
  mcp_tools:
    _type: mcp_client
    server:
      transport: streamable-http
      url: "http://localhost:9901/mcp"
    include:
      - tool_a
      - tool_b
    tool_overrides:
      tool_a:
        alias: "tool_a_alias"
        description: "Tool A description"

workflow:
  _type: react_agent
  tool_names:
    - mcp_tools

You can use the mcp_client function group to connect to an MCP server, dynamically discover the tools it serves, and register them as NeMo Agent toolkit functions.

The function group supports filtering using the include and exclude parameters. You can also optionally override the tool name and description defined by the MCP server using the tool_overrides parameter.

The function group can be directly referenced in the workflow configuration and provides all accessible tools from the MCP server to the workflow. Multiple function groups can be used in the same workflow to access tools from multiple MCP servers. Refer to Function Groups for more information about function group capabilities.

A tool within a function group can also be referenced by its name using the following syntax: <function_group_name>.<tool_name>.

Note

This requires that the tool name is explicitly listed under the optional include list of the function group configuration.

See function group accessibility for more details.

Example:

workflows:
  _type: react_agent
  tool_names:
    - mcp_tools.tool_a

An additional case to note is when a function group is served by an MCP server, the tools within the function group must still be accessed by their full name. This is the same as the prior case, but there is an important difference. Consider the following example:

workflow:
  _type: react_agent
  tool_names:
    - mcp_tools__calculator__add

mcp_tools is the name of the function group, and calculator__add is the name of the tool within the function group. This is because the tools are added to the function group as functions, and the function group is then added to the workflow as a tool.

Configuration Options#

The mcp_client function group supports the following configuration options:

Note: You can get the complete list of configuration options and their schemas by running:

nat info components -t function_group -q mcp_client

Server Configuration#

server.transport: Transport type (stdio, sse, or streamable-http). Refer to Transport Configuration for details.
server.url: URL of the MCP server (required for sse and streamable-http transports)
server.command: Command to run for stdio transport, such as python or docker
server.args: Arguments for the stdio command
server.env: Environment variables for the stdio process
server.auth_provider: Reference to authentication provider for protected MCP servers (only supported with streamable-http transport)

Timeout Configuration#

tool_call_timeout: Timeout for MCP tool calls. Defaults to 60 seconds
auth_flow_timeout: Timeout for interactive authentication flow. Defaults to 300 seconds

Reconnection Configuration#

reconnect_enabled: Whether to enable reconnecting to the MCP server if the connection is lost. Defaults to true.
reconnect_max_attempts: Maximum number of reconnect attempts. Defaults to 2.
reconnect_initial_backoff: Initial backoff time for reconnect attempts. Defaults to 0.5 seconds.
reconnect_max_backoff: Maximum backoff time for reconnect attempts. Defaults to 50.0 seconds.

Session Management Configuration#

max_sessions: Maximum number of concurrent session clients. Defaults to 100.
session_idle_timeout: Time after which inactive sessions are cleaned up. Defaults to 1 hour.

Tool Customization#

tool_overrides: Optional overrides for tool names and descriptions. Each entry can specify:
- alias: Override the tool name (function name in the workflow)
- description: Override the tool description

Example with all options:

function_groups:
  mcp_tools:
    _type: mcp_client
    include:
      - calculator__add
      - calculator__multiply
    server:
      transport: streamable-http
      url: "http://localhost:9901/mcp"
      auth_provider: "mcp_oauth2"  # Optional authentication
    tool_call_timeout: 60  # 1 minute for tool calls
    auth_flow_timeout: 300  # 5 minutes for auth flow
    reconnect_enabled: true
    reconnect_max_attempts: 3
    reconnect_initial_backoff: 1.0
    reconnect_max_backoff: 60.0
    max_sessions: 50  # Maximum concurrent sessions
    session_idle_timeout: 7200  # 2 hours (in seconds)
    tool_overrides:
      calculator__add:
        alias: "add_numbers"
        description: "Add two numbers together"
      calculator__multiply:
        description: "Multiply two numbers"  # Keeps original name

Transport Configuration#

The mcp_client function group can connect to MCP servers using different transport types. Choose the transport that matches your MCP server’s configuration to ensure proper communication.

Transport Types#

streamable-http (default): Modern HTTP-based transport, recommended for new deployments
SSE: Server-Sent Events transport, maintained for backwards compatibility
stdio: Standard input/output transport for local process communication

Streamable-HTTP Mode Configuration#

For streamable-http mode, you only need to specify the server URL:

functions:
  mcp_client:
    _type: mcp_client
    server:
      transport: streamable-http
      url: "http://localhost:8080/mcp"

SSE Mode Configuration#

SSE mode is supported for backward compatibility with existing systems. It is recommended to use streamable-http mode instead.

function_groups:
  mcp_tools:
    _type: mcp_client
    server:
      transport: sse
      url: "http://localhost:8080/sse"

STDIO Mode Configuration#

For STDIO mode, you need to specify the command to run and any additional arguments or environment variables:

function_groups:
  github_mcp:
    _type: mcp_client
    server:
      transport: stdio
      command: "docker"
      args: [
        "run",
        "-i",
        "--rm",
        "-e",
        "GITHUB_PERSONAL_ACCESS_TOKEN",
        "ghcr.io/github/github-mcp-server"
      ]
      env:
        GITHUB_PERSONAL_ACCESS_TOKEN: "${input:github_token}"

Example#

The following example demonstrates how to use the mcp_client function group with both local and remote MCP servers. This configuration shows how to use multiple MCP servers with different transports in the same workflow.

examples/MCP/simple_calculator_mcp/configs/config-mcp-client.yml:

function_groups:
  mcp_time:
    _type: mcp_client
    server:
      transport: stdio
      command: "python"
      args: ["-m", "mcp_server_time", "--local-timezone=America/Los_Angeles"]
  mcp_math:
    _type: mcp_client
    server:
      transport: streamable-http
      url: "http://localhost:9901/mcp"

workflow:
  _type: react_agent
  tool_names:
    - mcp_time
    - mcp_math

This configuration creates two function groups:

mcp_time: Connects to a local MCP server using stdio transport to get current date and time
mcp_math: Connects to a remote MCP server using streamable-http transport to access calculator tools

To run this example:

Start the remote MCP server:

nat mcp serve --config_file examples/getting_started/simple_calculator/configs/config.yml

This starts an MCP server on port 9901 with endpoint /mcp and uses streamable-http transport. Refer to MCP Server for more information.

Run the workflow:

nat run --config_file examples/MCP/simple_calculator_mcp/configs/config-mcp-client.yml --input "Is the product of 2 * 4 greater than the current hour of the day?"

Displaying MCP Tools using the CLI#

Use the nat mcp client commands to inspect and call tools available from an MCP server before configuring your workflow. This is useful for discovering available tools and understanding their input schemas.

List All Tools#

To list all tools served by an MCP server:

# For streamable-http transport (default)
nat mcp client tool list --url http://localhost:9901/mcp

# For stdio transport
nat mcp client tool list --transport stdio --command "python" --args "-m mcp_server_time"

# For SSE transport
nat mcp client tool list --url http://localhost:9901/sse --transport sse

For SSE transport, ensure the MCP server starts with the --transport sse flag. The transport type on the client and server needs to match for MCP communication to work. The default transport type is streamable-http.

Sample output:

calculator__add
calculator__multiply
calculator__subtract
calculator__divide
calculator__compare
current_datetime
react_agent

Get Tool Details#

To get detailed information about a specific tool, use the --tool flag:

nat mcp client tool list --url http://localhost:9901/mcp --tool calculator__multiply

Sample output:

Tool: calculator__multiply
Description: Multiply two or more numbers together
Input Schema:
{
  "properties": {
    "numbers": {
      "description": "",
      "items": {
        "type": "number"
      },
      "title": "Numbers",
      "type": "array"
    }
  },
  "required": [
    "numbers"
  ],
  "title": "Calculator__MultiplyInputSchema",
  "type": "object"
}

Call a Tool#

To call a tool and get its output:

# Pass arguments as JSON
$ nat mcp client tool call calculator__multiply \
  --url http://localhost:9901/mcp \
  --json-args '{"numbers": [1, 3, 6, 10]}'
180.0

Using Protected MCP Servers#

To use a protected MCP server, you need to provide the --auth flag:

nat mcp client tool list --url http://example.com/mcp --auth

This will use the mcp_oauth2 authentication provider to authenticate the user. For more information, refer to MCP Authentication.

List MCP Client Tools using the HTTP endpoint#

This is useful when you want to inspect the tools configured on the client side and whether each tool is available on the connected server.

When you serve a workflow that includes an mcp_client function group, the NeMo Agent toolkit exposes an HTTP endpoint to inspect the tools configured on the client side and whether each tool is available on the connected server.

Steps#

Start the MCP server:

nat mcp serve --config_file examples/getting_started/simple_calculator/configs/config.yml

Start the workflow (MCP client) with FastAPI:

nat serve --config_file examples/MCP/simple_calculator_mcp/configs/config-mcp-client.yml

Call the endpoint and pretty-print the response:

curl -s http://localhost:8000/mcp/client/tool/list | jq

Endpoint#

Path: /mcp/client/tool/list
Method: GET
Purpose: Returns tools configured in each mcp_client function group, indicates whether each tool is available on the connected MCP server, and includes metadata about the function group and HTTP session.

Sample Output#

{
  "mcp_clients": [
    {
      "function_group": "mcp_time",
      "server": "stdio:python",
      "transport": "stdio",
      "session_healthy": true,
      "protected": false,
      "tools": [
        {
          "name": "convert_time",
          "description": "Convert time between timezones",
          "server": "stdio:python",
          "available": true
        },
        {
          "name": "get_current_time_mcp_tool",
          "description": "Returns the current date and time",
          "server": "stdio:python",
          "available": true
        }
      ],
      "total_tools": 2,
      "available_tools": 2
    },
    {
      "function_group": "mcp_math",
      "server": "streamable-http:http://localhost:9901/mcp",
      "transport": "streamable-http",
      "session_healthy": true,
      "protected": false,
      "tools": [
        {
          "name": "calculator__add",
          "description": "Add two or more numbers together",
          "server": "streamable-http:http://localhost:9901/mcp",
          "available": true
        },
        {
          "name": "calculator__compare",
          "description": "Compare two numbers",
          "server": "streamable-http:http://localhost:9901/mcp",
          "available": true
        },
        {
          "name": "calculator__divide",
          "description": "Divide one number by another",
          "server": "streamable-http:http://localhost:9901/mcp",
          "available": true
        },
        {
          "name": "calculator__multiply",
          "description": "Multiply two or more numbers together",
          "server": "streamable-http:http://localhost:9901/mcp",
          "available": true
        },
        {
          "name": "calculator__subtract",
          "description": "Subtract one number from another",
          "server": "streamable-http:http://localhost:9901/mcp",
          "available": true
        }
      ],
      "total_tools": 5,
      "available_tools": 5
    }
  ]
}

MCP Inspection via UI#

You can inspect the MCP tools available on the client side using the UI.

Steps#

Start the MCP server:

nat mcp serve --config_file examples/getting_started/simple_calculator/configs/config.yml

Start the workflow (MCP client) with FastAPI:

nat serve --config_file examples/MCP/simple_calculator_mcp/configs/config-mcp-client.yml

Launch the UI by following the instructions in the Launching the UI documentation.
Click on the MCP tab in the side panel to inspect the MCP tools available on the client side.

Sample Output#

MCP Side Panel

MCP Tools

Troubleshooting#

If you encounter connection issues:

Verify the MCP server is running and accessible using the example nat mcp client ping command:
```
nat mcp client ping --url http://localhost:9901/mcp
```
Check that the transport type matches the server configuration
Ensure the URL or command is correct
Check network connectivity for remote servers

NVIDIA NeMo Agent Toolkit as an MCP Client#

Installation#

Accessing Protected MCP Servers#

MCP Client Configuration#

mcp_client Configuration#

Configuration Options#

Server Configuration#

Timeout Configuration#

Reconnection Configuration#

Session Management Configuration#

Tool Customization#

Transport Configuration#

Transport Types#

Streamable-HTTP Mode Configuration#

SSE Mode Configuration#

STDIO Mode Configuration#

Example#

Displaying MCP Tools using the CLI#

List All Tools#

Get Tool Details#

Call a Tool#

Using Protected MCP Servers#

List MCP Client Tools using the HTTP endpoint#

Steps#

Endpoint#

Sample Output#

MCP Inspection via UI#

Steps#

Sample Output#

Troubleshooting#

`mcp_client` Configuration#