Provisioning Machines using MCPs

Source: fly.io Date: 2025-05-12 URL: https://fly.io/blog/mcp-provisioning/

Summary

Engineering writeup with visionary elements documenting early MCP integration into flyctl, enabling LLMs like Claude to manage Fly infrastructure (volumes, apps, machines, logs) through natural language conversation. The post shows a working prototype — Claude creating and deleting Fly volumes via MCP — while acknowledging most commands are “roughed in” and full integration is weeks away. The framing is honest: “some assembly required,” but the direction is clear.

Implications

MCP integration / Machines API as agent-runtime substrate. This is an early entry in the “Sprites speak MCP” thread — Fly.io treating MCP as the primary interface between LLM agents and their infrastructure. The significance isn’t the prototype itself but the commitment: flyctl as an MCP server positions Fly as the infrastructure layer that AI agents natively manage, not just consume. This is competitive pressure on Vercel and Render, whose CLIs have no equivalent capability. The deeper question is whether MCP-as-infra-control-plane becomes standard — if so, every cloud provider needs an MCP endpoint, and fly.io’s head start compounds.