stand-in

A stand-in for your MCP server boilerplate.

You write with stand-in declarative macros that look like your MCP server — tools, resources, prompts — but when the compiler rolls, the macros step aside and production-ready code takes their place. You never touch the generated code. You only ever interact with the stand-in.

Status

🚧 Early Development — Core macros (#[mcp_tool], #[mcp_server], #[mcp_prompt], #[mcp_resource]) and both transports (Stdio, Streamable HTTP) are implemented. APIs may change before 1.0.

Installation

Add to your Cargo.toml:

[dependencies]

stand-in = "0.0.4"

tokio = { version = "1", features = ["rt-multi-thread", "macros"] }

Feature Flags

To enable HTTP transport:

stand-in = { version = "0.0.3", features = ["http"] }

Quick Start

use stand_in::prelude::*;

#[mcp_tool(
    name = "get_weather",
    description = "Returns current weather for a given city"
)]
async fn get_weather(city: String) -> Result<String> {
    let forecast = fetch_weather(&city).await?;
    Ok(format!("{}: {}°C, {}", city, forecast.temp, forecast.condition))
}

#[mcp_server]
struct MyServer;

#[tokio::main]
async fn main() {
    MyServer::serve(StdioTransport::default()).await;
}

That's it. No handler registration. No JSON-RPC routing. No protocol plumbing. The stand-in handles the setup; the compiler delivers the performance.

Adding a Prompt

use stand_in::prelude::*;

#[mcp_prompt(
    name = "summarize",
    description = "Summarize a document for a given audience"
)]
async fn summarize(document: String, audience: Option<String>) -> Result<Prompt> {
    let level = audience.as_deref().unwrap_or("general");
    Ok(Prompt::user(format!(
        "Summarize the following for a {level} audience:\n\n{document}"
    )))
}

Option<T> parameters become optional arguments in the MCP prompt definition. Required parameters stay required. The return type is always Result<Prompt>.

Adding a Resource

use stand_in::prelude::*;

/// Concrete resource — fixed URI, no parameters.
#[mcp_resource(
    uri = "info://version",
    name = "Server Version",
    mime_type = "application/json"
)]
async fn server_version() -> Result<String> {
    Ok(serde_json::json!({"version": "1.0.0"}).to_string())
}

/// Template resource — URI template with {param}, extracted at read time.
#[mcp_resource(
    uri = "docs://{topic}/readme",
    name = "Documentation",
    description = "Documentation for a given topic",
    mime_type = "text/markdown"
)]
async fn docs_readme(topic: String) -> Result<String> {
    Ok(format!("# {topic}\n\nDocumentation for {topic}."))
}

Resources with {param} in the URI become template resources. Concrete resources (no {param}) appear in resources/list; templates appear in resources/templates/list. Return Result<Vec<u8>> for binary data — the macro auto-detects the return type and produces base64-encoded BlobResourceContents.

Philosophy

Inspired by frameworks like Spring Boot, stand-in follows a simple principle: convention eliminates configuration. If the shape of your code already tells us what you mean, you shouldn't have to say it twice.

The MCP protocol is well-defined but verbose to implement. Every server needs the same handshake, the same capability negotiation, the same JSON-RPC dispatch. stand-in absorbs all of that behind derive macros and attribute macros, so you focus on what your server does — not on how it speaks the protocol.

Workspace Structure

The project is organized as a Cargo workspace with two crates:

stand-in/
├── Cargo.toml              # workspace root
├── stand-in/
│   ├── Cargo.toml
│   └── src/
│       ├── lib.rs           # re-exports, runtime, transports
│       └── ...
└── stand-in-macros/
    ├── Cargo.toml
    └── src/
        └── lib.rs           # proc macros: #[mcp_server], #[mcp_tool], etc.

Features

• #[mcp_tool] — Declare a tool with typed parameters. Schema is inferred from the function signature.
• #[mcp_prompt] — Define reusable prompt templates with typed arguments. Arguments are inferred from the function signature; Option<T> parameters are optional.
• #[mcp_server] — Wire everything together. Generates initialization, capability negotiation, and dispatch.
• Transports — Stdio (default) and Streamable HTTP (feature-gated). Extensible via the Transport trait.
• Async-first — Built on tokio. Every handler is async fn.
• #[mcp_resource] — Expose data as MCP resources (concrete URIs or URI templates with {param}). Return Result<String> for text content or Result<Vec<u8>> for base64-encoded blobs.

Example: A More Complete Server

use stand_in::prelude::*;

/// A server that exposes project management tools and prompts.
/// Server name and version are read from Cargo.toml at compile time.
#[mcp_server]
struct ProjectHub;

#[mcp_tool(name = "list_tasks", description = "List all open tasks for a project")]
async fn list_tasks(project_id: String) -> Result<String> {
    // ... fetch from database
    Ok(format!("Tasks for {project_id}: ..."))
}

#[mcp_tool(name = "create_task", description = "Create a new task")]
async fn create_task(project_id: String, title: String, assignee: Option<String>) -> Result<String> {
    // ... write to database
    Ok(format!("Created task '{title}' in {project_id}"))
}

#[mcp_prompt(
    name = "summarize_project",
    description = "Generate a project status summary"
)]
async fn summarize_project(project_id: String, format: Option<String>) -> Result<Prompt> {
    let level = format.as_deref().unwrap_or("brief");
    Ok(Prompt::user(format!(
        "Summarize project {project_id} in a {level} format."
    )))
}

#[mcp_resource(
    uri = "project://{project_id}/overview",
    name = "Project Overview",
    mime_type = "application/json"
)]
async fn project_overview(project_id: String) -> Result<String> {
    Ok(serde_json::json!({"id": project_id, "status": "active"}).to_string())
}

#[tokio::main]
async fn main() {
    ProjectHub::serve(StdioTransport::default()).await.unwrap();
}

Why "stand-in"?

Because good infrastructure disappears.

A stand-in does essential work — they're on set for hours so the real performance can happen in minutes. But you never see them in the final cut. That's exactly what these macros do: they show up in your source code, do the hard work at compile time, and vanish from the binary.

Your code reads like a declaration of intent. The compiler turns it into a server. The stand-in was never in the final cut.

License

MIT