Crate llm_tool_mcp 

llm-tool-mcp

MCP (Model Context Protocol) stdio server for llm-tool registries.

Register your tools in a ToolRegistry, hand it to McpServer, and get a fully compliant MCP server — no boilerplate.

Quick start

use llm_tool::{llm_prompt, llm_resource, llm_tool, ToolContext, ToolError, ToolRegistry};
use llm_tool_mcp::McpServer;

/// Adds two numbers.
#[llm_tool]
fn add(
    /// First operand.
    a: i64,
    /// Second operand.
    b: i64,
) -> Result<String, ToolError> {
    Ok(format!("{}", a + b))
}

/// Code review instruction template.
#[llm_prompt]
fn review_prompt(
    /// Programming language.
    lang: String,
) -> String {
    format!("Please review this {lang} code for security bugs.")
}

/// Dynamic application config resource.
#[llm_resource(uri = "file:///config/{app}.json")]
fn get_config(app: String) -> String {
    format!(r#"{{"app":"{app}","enabled":true}}"#)
}

let registry = ToolRegistry::new().with_tool(Add);

let server = McpServer::builder("my-server", "0.1.0", registry)
    .with_prompt(ReviewPrompt)
    .with_resource(GetConfig)
    .with_context(ToolContext::new().with_conversation_id("caller-id"))
    .build();

// In production: server.run_stdio().expect("server failed");
// Here we feed a request via an in-memory buffer:
let input = r#"{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{"name":"add","arguments":{"a":17,"b":25}}}"#;
let reader = std::io::Cursor::new(format!("{input}\n"));
let mut output = Vec::new();
server.run(reader, &mut output).unwrap();

let resp: serde_json::Value = serde_json::from_slice(&output).unwrap();
assert_eq!(resp["result"]["content"][0]["text"], "42");

Transports: Stdio vs TCP vs Unix Sockets

McpServer is builder-style and supports all common execution models out-of-the-box. Two flavors are available:

• Blocking convenience — run_stdio, run_tcp, run_unix, and the transport-dispatching serve(Transport). These build a Tokio runtime internally, so a simple binary’s main needs no async. They block until the transport finishes.
• Async first-class — run_async, listen_tcp, listen_unix. Use these when you already have a Tokio runtime (as most real applications do) so the server shares it instead of spawning a second one.

let server = McpServer::new("my-server", "0.1.0", ToolRegistry::new());

// ── Blocking convenience (no async main required) ──

// 1. Standard MCP desktop client transport (stdio subprocess):
// server.run_stdio().expect("stdio server failed");

// 2. TCP network server (localhost only):
// server.run_tcp("127.0.0.1:3000").expect("tcp server failed");

// 3. Unix Domain Socket (local IPC):
// server.run_unix("/tmp/my-agent.sock").expect("unix server failed");

// 4. Pick a transport at runtime (e.g. from CLI flags) and serve it:
let transport = Transport::Tcp("0.0.0.0:8080".parse().unwrap());
// server.serve(transport).expect("server failed");

let server = McpServer::new("my-server", "0.1.0", ToolRegistry::new());

// ── Async first-class (inside an existing Tokio runtime) ──

// server.listen_tcp("127.0.0.1:3000").await.expect("tcp bind failed");
// server.listen_unix("/tmp/my-agent.sock").await.expect("unix bind failed");

What it handles

MCP method	Behavior
initialize	Returns server info and capabilities for registered primitives
notifications/initialized	Acknowledged silently
tools/list	Derives schemas from ToolRegistry::definitions()
tools/call	Dispatches via ToolRegistry::dispatch(), returns content
prompts/list	Lists all registered prompts and their argument schemas
prompts/get	Renders prompt messages with argument substitution
resources/list	Lists all static resources registered on the server
resources/templates/list	Lists all URI templates (e.g. "file:///config/{app}.json")
resources/read	Matches URIs against resources/templates and returns content

Tool errors are returned as MCP content with isError: true (spec-compliant), not as JSON-RPC errors.

Async & custom transports

If running inside an existing Tokio application or network server, use run_async:

let server = McpServer::new("s", "1", ToolRegistry::new());

// Runs over any tokio::io::AsyncBufRead + AsyncWrite streams:
// server.run_async(tokio::io::stdin(), tokio::io::stdout()).await.unwrap();

For custom request/response routing (e.g. Axum HTTP POST or WebSockets), call handle_message. It accepts a single request or a JSON-RPC batch array and returns a structured RpcOutcome — a Single response object or a Batch array — which you can inspect or render to the wire in a single pass with .to_wire(). None means the input was purely a notification, so there is nothing to send back:

let server = McpServer::new("s", "1", ToolRegistry::new());

let request = r#"{"jsonrpc":"2.0","id":1,"method":"tools/list"}"#;
match server.handle_message(request).await {
    // `outcome` is a `Single` object or a `Batch` array — render it directly.
    Some(outcome) => {
        let body = outcome.to_wire();
        // ...write `body` to your HTTP/WebSocket response...
        assert!(body.contains("\"result\""));
    }
    // Notification-only input: reply 202/204 with no body.
    None => {}
}

License

Dual-licensed under Apache-2.0 OR MIT.

Modules

protocol

JSON-RPC 2.0 protocol types for MCP communication.

Structs

Connection

Per-connection negotiated state, owned by each transport run loop.

McpServer

An MCP server that serves tools from a ToolRegistry over JSON-RPC.

McpServerBuilder

Builder for McpServer that registers prompts and resources up front.

Enums

RpcOutcome

A dispatched JSON-RPC result, ready to be inspected or rendered.

Transport

A transport for the blocking McpServer::serve entry point.

Traits

RegistryFactory

Builds the ToolRegistry a given caller should see.