RailwayMCP Server

Railway's official MCP server gives an agent operational control of a Railway account: the projects, services and environments in it, the deployments running, the variables and domains configured, and the logs and metrics coming out. It is distributed with the Railway CLI rather than as a standalone package — the earlier @railway/mcp-server npm build is deprecated in favour of railway mcp.

Local and remote are different servers

This is the detail worth knowing before wiring it up.

Local MCP runs through the CLI on your machine, reusing the credentials from railway login. It is the full surface: roughly 40 tools spanning account, projects, services, environments, deployments, variables, domains, networking, templates and storage, plus observability tools for logs, metrics, request analysis and error rates.

Remote MCP runs at mcp.railway.com and exposes a deliberately narrower set — account, projects, feature flags and deployments — plus a railway-agent tool that handles multi-step work like log analysis, debugging and service configuration on the server side. It authenticates by direct OAuth, or through the CLI acting as a proxy for your existing login. Project tokens are not accepted; remote MCP requires a user identity.

Safety model

Destructive operations carry protocol-level hints and require an explicit confirm: true before they execute, so a deletion cannot fall out of a single ambiguous instruction. Remote MCP issues short-lived OAuth tokens and marks destructive actions for client-side confirmation. Railway's own guidance is to review anything an LLM proposes against a production environment before approving it.

Setup

railway setup agent configures the local server; --remote switches to the CLI-proxied remote server and --remote --oauth to a direct OAuth connection. railway mcp install does per-editor configuration by hand. Cursor, VS Code, Claude Code and Windsurf are documented as supported clients.