railwayapp 5.30.3

Interact with Railway via CLI
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56

use super::*;
use rmcp::{ServiceExt, transport::stdio};

mod handler;
pub(crate) mod install;
pub(crate) mod params;
mod proxy;
mod tools;
use handler::RailwayMcp;

/// Starts a local MCP server for AI-agent access, or installs the MCP config into AI coding tools.
#[derive(Parser)]
pub struct Args {
    #[clap(subcommand)]
    command: Option<Commands>,
}

#[derive(Parser)]
enum Commands {
    /// Install Railway's MCP server config into AI coding tools (Claude Code, Cursor, OpenCode, Codex)
    Install(install::Args),
    /// Proxy the remote MCP server (mcp.railway.com) over stdio, authenticating with your CLI login
    Proxy,
}

pub async fn command(args: Args) -> Result<()> {
    match args.command {
        None => serve_stdio().await,
        Some(Commands::Install(install_args)) => install::command(install_args).await,
        Some(Commands::Proxy) => proxy::serve_proxy().await,
    }
}

async fn serve_stdio() -> Result<()> {
    let configs = Configs::new()?;
    // Start even when there are no usable credentials. Refusing to boot makes
    // the harness report an opaque "MCP server failed" that no later
    // `railway login` can clear, because the process is already gone. Serving
    // with an unauthenticated client instead means tool calls return an
    // actionable auth error, and `refresh_credentials` swaps in an authorized
    // client as soon as the user signs in — no editor restart.
    let client = match GQLClient::new_authorized(&configs) {
        Ok(client) => client,
        Err(_) => GQLClient::new_public()?,
    };
    let handler = RailwayMcp::new(client, configs);

    let service = handler
        .serve(stdio())
        .await
        .context("Failed to start MCP server")?;

    service.waiting().await?;

    Ok(())
}