MCP Tools

Model Context Protocol (MCP) is an open standard that enables LLMs to communicate with external applications, data sources, and tools. ADK-Rust provides full MCP support through the McpToolset, allowing you to connect to any MCP-compliant server and expose its tools to your agents.

Overview

MCP follows a client-server architecture:

• MCP Servers expose tools, resources, and prompts
• MCP Clients (like ADK agents) connect to servers and use their capabilities

Benefits of MCP integration:

• Universal connectivity - Connect to any MCP-compliant server
• Automatic discovery - Tools are discovered dynamically from the server
• Language agnostic - Use tools written in any language
• Growing ecosystem - Access thousands of existing MCP servers

Prerequisites

MCP servers are typically distributed as npm packages. You'll need:

• Node.js and npm installed
• An LLM API key (Gemini, OpenAI, etc.)

Quick Start

Connect to an MCP server and use its tools:

use adk_agent::LlmAgentBuilder;
use adk_core::{Content, Part, ReadonlyContext, Toolset};
use adk_model::GeminiModel;
use adk_tool::McpToolset;
use rmcp::{ServiceExt, transport::TokioChildProcess};
use tokio::process::Command;
use std::sync::Arc;

#[tokio::main]
async fn main() -> anyhow::Result<()> {
    dotenvy::dotenv().ok();
    let api_key = std::env::var("GOOGLE_API_KEY")?;
    let model = Arc::new(GeminiModel::new(&api_key, "gemini-2.0-flash")?);

    // 1. Start MCP server and connect
    let mut cmd = Command::new("npx");
    cmd.arg("-y").arg("@modelcontextprotocol/server-everything");

    let client = ().serve(TokioChildProcess::new(cmd)?).await?;

    // 2. Create toolset from the client
    let toolset = McpToolset::new(client)
        .with_tools(&["echo", "add"]);  // Only expose these tools

    // 3. Get cancellation token for cleanup
    let cancel_token = toolset.cancellation_token().await;

    // 4. Discover tools and add to agent
    let ctx: Arc<dyn ReadonlyContext> = Arc::new(SimpleContext);
    let tools = toolset.tools(ctx).await?;

    let mut builder = LlmAgentBuilder::new("mcp_agent")
        .model(model)
        .instruction("You have MCP tools. Use 'echo' to repeat messages, 'add' to sum numbers.");

    for tool in tools {
        builder = builder.tool(tool);
    }

    let agent = builder.build()?;

    // 5. Run interactive console
    adk_cli::console::run_console(
        Arc::new(agent),
        "mcp_demo".to_string(),
        "user".to_string(),
    ).await?;

    // 6. Cleanup: shutdown MCP server
    cancel_token.cancel();

    Ok(())
}

// Minimal context for tool discovery
struct SimpleContext;

#[async_trait::async_trait]
impl ReadonlyContext for SimpleContext {
    fn invocation_id(&self) -> &str { "init" }
    fn agent_name(&self) -> &str { "init" }
    fn user_id(&self) -> &str { "user" }
    fn app_name(&self) -> &str { "mcp" }
    fn session_id(&self) -> &str { "init" }
    fn branch(&self) -> &str { "main" }
    fn user_content(&self) -> &Content {
        static CONTENT: std::sync::OnceLock<Content> = std::sync::OnceLock::new();
        CONTENT.get_or_init(|| Content::new("user").with_text("init"))
    }
}

Run with:

GOOGLE_API_KEY=your_key cargo run --bin basic

McpToolset API

Creating a Toolset

use adk_tool::McpToolset;

// Basic creation
let toolset = McpToolset::new(client);

// With custom name
let toolset = McpToolset::new(client)
    .with_name("filesystem-tools");

Tool Filtering

Filter which tools to expose:

// Filter by predicate function
let toolset = McpToolset::new(client)
    .with_filter(|name| {
        matches!(name, "read_file" | "write_file" | "list_directory")
    });

// Filter by exact names (convenience method)
let toolset = McpToolset::new(client)
    .with_tools(&["echo", "add", "get_time"]);

Cleanup with Cancellation Token

Always get a cancellation token to cleanly shutdown the MCP server:

let toolset = McpToolset::new(client);
let cancel_token = toolset.cancellation_token().await;

// ... use the toolset ...

// Before exiting, shutdown the MCP server
cancel_token.cancel();

This prevents EPIPE errors and ensures clean process termination.

Connecting to MCP Servers

Local Servers (Stdio)

Connect to a local MCP server via standard input/output:

use rmcp::{ServiceExt, transport::TokioChildProcess};
use tokio::process::Command;

// NPM package server
let mut cmd = Command::new("npx");
cmd.arg("-y")
    .arg("@modelcontextprotocol/server-filesystem")
    .arg("/path/to/allowed/directory");
let client = ().serve(TokioChildProcess::new(cmd)?).await?;

// Local binary server
let mut cmd = Command::new("./my-mcp-server");
cmd.arg("--config").arg("config.json");
let client = ().serve(TokioChildProcess::new(cmd)?).await?;

Remote Servers (SSE)

Connect to a remote MCP server via Server-Sent Events:

use rmcp::{ServiceExt, transport::SseClient};

let client = ().serve(
    SseClient::new("http://localhost:8080/sse")?
).await?;

Tool Discovery

The McpToolset automatically discovers tools from the connected server:

use adk_core::{ReadonlyContext, Toolset};

// Get discovered tools
let tools = toolset.tools(ctx).await?;

println!("Discovered {} tools:", tools.len());
for tool in &tools {
    println!("  - {}: {}", tool.name(), tool.description());
}

Each discovered tool:

• Has its name and description from the MCP server
• Includes parameter schemas for LLM accuracy
• Executes via the MCP protocol when called

Adding Tools to Agent

There are two patterns for adding MCP tools to an agent:

Pattern 1: Add as Toolset

let toolset = McpToolset::new(client);

let agent = LlmAgentBuilder::new("agent")
    .model(model)
    .toolset(Arc::new(toolset))
    .build()?;

Pattern 2: Add Individual Tools

This gives you more control over which tools are added:

let toolset = McpToolset::new(client)
    .with_tools(&["echo", "add"]);

let tools = toolset.tools(ctx).await?;

let mut builder = LlmAgentBuilder::new("agent")
    .model(model);

for tool in tools {
    builder = builder.tool(tool);
}

let agent = builder.build()?;

Popular MCP Servers

Here are some commonly used MCP servers you can integrate:

Everything Server (Testing)

npx -y @modelcontextprotocol/server-everything

Tools: echo, add, longRunningOperation, sampleLLM, getAlerts, printEnv

Filesystem Server

npx -y @modelcontextprotocol/server-filesystem /path/to/directory

Tools: read_file, write_file, list_directory, search_files

GitHub Server

npx -y @modelcontextprotocol/server-github

Tools: search_repositories, get_file_contents, create_issue

Slack Server

npx -y @modelcontextprotocol/server-slack

Tools: send_message, list_channels, search_messages

Memory Server

npx -y @modelcontextprotocol/server-memory

Tools: store, retrieve, search

Find more servers at the MCP Server Registry.

Error Handling

Handle MCP connection and execution errors:

use adk_core::AdkError;

match toolset.tools(ctx).await {
    Ok(tools) => {
        println!("Discovered {} tools", tools.len());
    }
    Err(AdkError::Tool(msg)) => {
        eprintln!("MCP error: {}", msg);
    }
    Err(e) => {
        eprintln!("Other error: {}", e);
    }
}

Common errors:

• Connection failed - Server not running or wrong address
• Tool execution failed - MCP server returned an error
• Invalid parameters - Tool received incorrect arguments

Best Practices

1. Filter tools - Only expose tools the agent needs to reduce confusion
2. Use cancellation tokens - Always call cancel() before exiting to cleanup
3. Handle errors - MCP servers may fail; implement appropriate error handling
4. Use local servers - For development, stdio transport is simpler than remote
5. Check server status - Verify MCP server is running before creating toolset

Complete Example

Here's a full working example with proper cleanup:

use adk_agent::LlmAgentBuilder;
use adk_core::{Content, Part, ReadonlyContext, Toolset};
use adk_model::GeminiModel;
use adk_tool::McpToolset;
use rmcp::{ServiceExt, transport::TokioChildProcess};
use std::sync::Arc;
use tokio::process::Command;

struct SimpleContext;

#[async_trait::async_trait]
impl ReadonlyContext for SimpleContext {
    fn invocation_id(&self) -> &str { "init" }
    fn agent_name(&self) -> &str { "init" }
    fn user_id(&self) -> &str { "user" }
    fn app_name(&self) -> &str { "mcp" }
    fn session_id(&self) -> &str { "init" }
    fn branch(&self) -> &str { "main" }
    fn user_content(&self) -> &Content {
        static CONTENT: std::sync::OnceLock<Content> = std::sync::OnceLock::new();
        CONTENT.get_or_init(|| Content::new("user").with_text("init"))
    }
}

#[tokio::main]
async fn main() -> anyhow::Result<()> {
    dotenvy::dotenv().ok();

    let api_key = std::env::var("GOOGLE_API_KEY")?;
    let model = Arc::new(GeminiModel::new(&api_key, "gemini-2.0-flash")?);

    println!("Starting MCP server...");
    let mut cmd = Command::new("npx");
    cmd.arg("-y").arg("@modelcontextprotocol/server-everything");

    let client = ().serve(TokioChildProcess::new(cmd)?).await?;
    println!("MCP server connected!");

    // Create filtered toolset
    let toolset = McpToolset::new(client)
        .with_name("everything-tools")
        .with_filter(|name| matches!(name, "echo" | "add" | "printEnv"));

    // Get cancellation token for cleanup
    let cancel_token = toolset.cancellation_token().await;

    // Discover tools
    let ctx = Arc::new(SimpleContext) as Arc<dyn ReadonlyContext>;
    let tools = toolset.tools(ctx).await?;

    println!("Discovered {} tools:", tools.len());
    for tool in &tools {
        println!("  - {}: {}", tool.name(), tool.description());
    }

    // Build agent with tools
    let mut builder = LlmAgentBuilder::new("mcp_demo")
        .model(model)
        .instruction(
            "You have access to MCP tools:\n\
             - echo: Repeat a message back\n\
             - add: Add two numbers (a + b)\n\
             - printEnv: Print environment variables"
        );

    for tool in tools {
        builder = builder.tool(tool);
    }

    let agent = builder.build()?;

    // Run interactive console
    let result = adk_cli::console::run_console(
        Arc::new(agent),
        "mcp_demo".to_string(),
        "user".to_string(),
    ).await;

    // Cleanup
    println!("\nShutting down MCP server...");
    cancel_token.cancel();
    tokio::time::sleep(tokio::time::Duration::from_millis(100)).await;

    result?;
    Ok(())
}

Advanced: Custom MCP Server

You can create your own MCP server in Rust using the rmcp SDK:

use rmcp::{tool, tool_router, handler::server::tool::ToolRouter, model::*};

#[derive(Clone)]
pub struct MyServer {
    tool_router: ToolRouter<Self>,
}

#[tool_router]
impl MyServer {
    fn new() -> Self {
        Self { tool_router: Self::tool_router() }
    }

    #[tool(description = "Add two numbers")]
    async fn add(&self, a: i32, b: i32) -> Result<CallToolResult, ErrorData> {
        Ok(CallToolResult::success(vec![Content::text((a + b).to_string())]))
    }

    #[tool(description = "Multiply two numbers")]
    async fn multiply(&self, a: i32, b: i32) -> Result<CallToolResult, ErrorData> {
        Ok(CallToolResult::success(vec![Content::text((a * b).to_string())]))
    }
}

See the rmcp documentation for complete server implementation details.

Related

• Function Tools - Creating custom tools in Rust
• Built-in Tools - Pre-built tools included with ADK
• LlmAgent - Adding tools to agents
• rmcp SDK - Official Rust MCP SDK
• MCP Specification - Protocol documentation

---

Previous: ← UI Tools | Next: Sessions →