MCP for AI agents

What it is

AI agents drive FlashVMM directly, through 21 MCP tools.

flash-mcp gives any MCP client the full FlashVM lifecycle. Agents create VMs, run code, copy files, fork, checkpoint and clean up on their own.

An agent jailed by flash agents keeps them from inside its sandbox, as 22 tools: the same set without the three host-wide ones (pools, host, controller), plus try_approaches, keep_approach, discard_approaches and undo_steps. See it run →

Why it matters

The best sandbox for AI is one the AI can operate. Every control a person has in the CLI, an agent has in MCP.

• Full parity. Run, exec, copy, sleep, wake, pools, GPU, host and controller, all as tools.
• Built for agents. Structured results, file changes per command, live progress.
• Fleet reach. The same tools place and watch VMs across your fleet.

Runs on

• Linux x86_64
• Linux arm64
• macOS
• Windows x64
• Windows arm64

How it works

Point any MCP client at flash-mcp and ask in plain English.

CLI

{ "mcpServers": { "flashvm": { "command": "flash-mcp" } } }

MCP

flash_cli(args: ["run", "--bundle", "BUNDLE", "--id", "agent-1", "--vcpus", "1", "--mem-mib", "256", "--detach"])
flash_exec(vm_id: "agent-1", argv: ["sh", "-c", "cat /etc/os-release"], changes: true)
flash_ls()
flash_cli(args: ["rm", "agent-1"])

API

POST /v1/flashvm/guest/execs {"argv": ["sh", "-c", "uname -a"]}

The result

An AI assistant starts, uses and removes a VM by itself.

With an 18 ms cold boot and a 1.2 ms claim from a warm pool, an agent gets a fresh sandbox before it finishes writing its next sentence.