What Changed

Fly.io now offers an MCP server at sprites.dev/mcp that enables AI agents to request and manage Sprites—disposable cloud computers with durable filesystems that cost virtually nothing when idle. Agents authenticated through Claude Desktop or compatible tools can now programmatically provision Sprites to handle computational tasks, development workflows, and testing scenarios.

How It Works

Users can authenticate their Fly.io organization to the MCP server, giving agents access to the Sprites API. Once connected, agents can respond to natural language commands like:

• "On a new Sprite, reproduce this bug from issues/913 and capture logs"
• "On 3 new Sprites, benchmark this function across different query libraries and test latency"
• "On a new Sprite, run a load generator against this endpoint for 60 seconds"

Guardrails and Safety

Recognizing the risks of agents autonomously provisioning infrastructure, Fly.io built in protective constraints:

• Creation caps: Limit the total number of Sprites an agent can provision
• Name prefixes: All agent-created Sprites receive a configurable prefix for easy identification and cleanup
• Organization scoping: Agents access only a single designated Fly.io organization

Implementation Philosophy

The team emphasizes that while MCP provides access, the preferred approach for agent-driven infrastructure is CLI tools and discoverable APIs rather than large context windows filled with tool descriptions. Modern AI models (Claude, Gemini, Codex) can efficiently drive CLI commands without MCP overhead. The MCP server exists primarily for agents that cannot execute shell commands.

Use Cases

Developers can now delegate repetitive computational tasks to agents: dependency updates with testing, multi-variant development environments, log analysis, dataset exploration, load testing, and webhook monitoring—all without manually provisioning infrastructure.