Crate rmcp 

RMCP: Rust Model Context Protocol

rmcp is the official Rust implementation of the Model Context Protocol (MCP), a protocol designed for AI assistants to communicate with other services. This library can be used to build both servers that expose capabilities to AI assistants and clients that interact with such servers.

Quick Start

Server Implementation

Creating a server with tools is simple using the #[tool] macro:

use rmcp::{
    ServerHandler, ServiceExt,
    handler::server::tool::ToolRouter,
    model::*,
    tool, tool_handler, tool_router,
    transport::stdio,
    ErrorData as McpError,
};
use std::sync::Arc;
use tokio::sync::Mutex;

#[derive(Clone)]
pub struct Counter {
    counter: Arc<Mutex<i32>>,
    tool_router: ToolRouter<Self>,
}

#[tool_router]
impl Counter {
    fn new() -> Self {
        Self {
            counter: Arc::new(Mutex::new(0)),
            tool_router: Self::tool_router(),
        }
    }

    #[tool(description = "Increment the counter by 1")]
    async fn increment(&self) -> Result<CallToolResult, McpError> {
        let mut counter = self.counter.lock().await;
        *counter += 1;
        Ok(CallToolResult::success(vec![Content::text(
            counter.to_string(),
        )]))
    }

    #[tool(description = "Get the current counter value")]
    async fn get(&self) -> Result<CallToolResult, McpError> {
        let counter = self.counter.lock().await;
        Ok(CallToolResult::success(vec![Content::text(
            counter.to_string(),
        )]))
    }
}

// Implement the server handler
#[tool_handler]
impl ServerHandler for Counter {
    fn get_info(&self) -> ServerInfo {
        ServerInfo {
            instructions: Some("A simple counter that tallies the number of times the increment tool has been used".into()),
            capabilities: ServerCapabilities::builder().enable_tools().build(),
            ..Default::default()
        }
    }
}

// Run the server
#[tokio::main]
async fn main() -> Result<(), Box<dyn std::error::Error>> {
    // Create and run the server with STDIO transport
    let service = Counter::new().serve(stdio()).await.inspect_err(|e| {
        println!("Error starting server: {}", e);
    })?;
    service.waiting().await?;
    Ok(())
}

Structured Output

Tools can return structured JSON data with schemas. Use the Json wrapper:

#[derive(Serialize, Deserialize, JsonSchema)]
struct CalculationRequest {
    a: i32,
    b: i32,
    operation: String,
}

#[derive(Serialize, Deserialize, JsonSchema)]
struct CalculationResult {
    result: i32,
    operation: String,
}

#[tool(name = "calculate", description = "Perform a calculation")]
async fn calculate(&self, params: Parameters<CalculationRequest>) -> Result<Json<CalculationResult>, String> {
    let result = match params.0.operation.as_str() {
        "add" => params.0.a + params.0.b,
        "multiply" => params.0.a * params.0.b,
        _ => return Err("Unknown operation".to_string()),
    };

    Ok(Json(CalculationResult { result, operation: params.0.operation }))
}

The #[tool] macro automatically generates an output schema from the CalculationResult type.

Tasks

RMCP implements the task lifecycle from SEP-1686 so long-running or asynchronous tool calls can be queued and polled safely.

• Create: set the task field on CallToolRequestParam to ask the server to enqueue the tool call. The response is a CreateTaskResult that includes the generated task.task_id.
• Inspect: use tasks/get (GetTaskInfoRequest) to retrieve metadata such as status, timestamps, TTL, and poll interval.
• Await results: call tasks/result (GetTaskResultRequest) to block until the task completes and receive either the final CallToolResult payload or a protocol error.
• Cancel: call tasks/cancel (CancelTaskRequest) to request termination of a running task.

To expose task support, enable the tasks capability when building ServerCapabilities. The #[task_handler] macro and OperationProcessor utility provide reference implementations for enqueuing, tracking, and collecting task results.

Client Implementation

Creating a client to interact with a server:

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

#[tokio::main]
async fn main() -> Result<(), Box<dyn std::error::Error>> {
    // Connect to a server running as a child process
    let service = ()
        .serve(TokioChildProcess::new(Command::new("uvx").configure(
            |cmd| {
                cmd.arg("mcp-server-git");
            },
        ))?)
        .await?;

    // Get server information
    let server_info = service.peer_info();
    println!("Connected to server: {server_info:#?}");

    // List available tools
    let tools = service.list_tools(Default::default()).await?;
    println!("Available tools: {tools:#?}");

    // Call a tool
    let result = service
        .call_tool(CallToolRequestParams {
            meta: None,
            name: "git_status".into(),
            arguments: serde_json::json!({ "repo_path": "." }).as_object().cloned(),
            task: None,
        })
        .await?;
    println!("Result: {result:#?}");

    // Gracefully close the connection
    service.cancel().await?;
    Ok(())
}

For more examples, see the examples directory in the repository.

Transport Options

RMCP supports multiple transport mechanisms, each suited for different use cases:

transport-async-rw

Low-level interface for asynchronous read/write operations. This is the foundation for many other transports.

transport-io

For working directly with I/O streams (tokio::io::AsyncRead and tokio::io::AsyncWrite).

transport-child-process

Run MCP servers as child processes and communicate via standard I/O.

Example:

use rmcp::transport::TokioChildProcess;
use tokio::process::Command;

let transport = TokioChildProcess::new(Command::new("mcp-server"))?;
let service = client.serve(transport).await?;

Access with peer interface when handling message

You can get the Peer struct from NotificationContext and RequestContext.

impl ServerHandler for Handler {
    async fn on_progress(
        &self,
        notification: ProgressNotificationParam,
        context: NotificationContext<RoleServer>,
    ) {
        let peer = context.peer;
        let _ = peer
            .notify_logging_message(LoggingMessageNotificationParam {
                level: LoggingLevel::Info,
                logger: None,
                data: serde_json::json!({
                    "message": format!("Progress: {}", notification.progress),
                }),
            })
            .await;
    }
}

Manage Multi Services

For many cases you need to manage several service in a collection, you can call into_dyn to convert services into the same type.

let service = service.into_dyn();

Feature Flags

RMCP uses feature flags to control which components are included:

• client: Enable client functionality
• server: Enable server functionality and the tool system
• macros: Enable the #[tool] macro (enabled by default)
• Transport-specific features:
  • transport-async-rw: Async read/write support
  • transport-io: I/O stream support
  • transport-child-process: Child process support
  • transport-streamable-http-client / transport-streamable-http-server: HTTP streaming (client agnostic, see StreamableHttpClientTransport for details)
    • transport-streamable-http-client-reqwest: a default reqwest implementation of the streamable http client
• auth: OAuth2 authentication support
• schemars: JSON Schema generation (for tool definitions)

Transports

• transport-io: Server stdio transport
• transport-child-process: Client stdio transport
• transport-streamable-http-server streamable http server transport
• transport-streamable-http-client streamable http client transport

Transport

The transport type must implement the Transport trait, which allows it to send messages concurrently and receive messages sequentially. There are 2 pairs of standard transport types:

transport	client	server
std IO	TokioChildProcess	stdio
streamable http	StreamableHttpClientTransport	StreamableHttpService

IntoTransport trait

IntoTransport is a helper trait that implicitly converts a type into a transport type.

These types automatically implement IntoTransport:

1. A type that implements both futures::Sink and futures::Stream, or a tuple (Tx, Rx) where Tx is futures::Sink and Rx is futures::Stream.
2. A type that implements both tokio::io::AsyncRead and tokio::io::AsyncWrite, or a tuple (R, W) where R is tokio::io::AsyncRead and W is tokio::io::AsyncWrite.
3. A type that implements the Worker trait.
4. A type that implements the Transport trait.

License

This project is licensed under the terms specified in the repository’s LICENSE file.

Re-exports

pub use error::ErrorData;

pub use handler::client::ClientHandler;client

pub use handler::server::ServerHandler;server

pub use handler::server::wrapper::Json;server

pub use service::Peer;client or server

pub use service::Service;client or server

pub use service::ServiceError;client or server

pub use service::ServiceExt;client or server

pub use service::RoleClient;client

pub use service::serve_client;client

pub use service::RoleServer;server

pub use service::serve_server;server

pub use schemars;macros and server

pub use serde;macros

pub use serde_json;macros

Modules

handler

model

Basic data types in MCP specification

serviceclient or server

task_manager

transport

Transport

Macros

const_string

elicit_safeelicitation

Macro to mark types as safe for elicitation by verifying they generate object schemas.

objectmacros

Use this macro just like serde_json::json!

pastemacros and server

Enums

RmcpError

This is an unified error type for the errors could be returned by the service.

Type Aliases

ErrorDeprecated

Attribute Macros

promptmacros and server

prompt

prompt_handlermacros and server

prompt_handler

prompt_routermacros and server

prompt_router

task_handlermacros and server

task_handler

toolmacros and server

tool

tool_handlermacros and server

tool_handler

tool_routermacros and server

tool_router