Starts the MCP server of fcvm, which gives an AI agent disposable Firecracker microVMs as sandboxes. Each sandbox boots its own kernel in about a second from any OCI image. The agent gets tools to run commands, read and write files, build images, commit prepared states, and snapshot a VM to fork running copies in about 150 ms. Network access is off, full, or an allowlist of hosts enforced by a logging proxy.

This package is only a launcher: it finds the installed fcvm and runs fcvm mcp. fcvm itself needs a Linux host with KVM:

curl -fsSL https://raw.githubusercontent.com/carlosbravoa/fcvm/main/install.sh | sh
fcvm setup

Then register the server with your MCP client. With Claude Code:

claude mcp add fcvm -- uvx fcvm-mcp
claude mcp add fcvm -e FCVM_MCP_NETWORK=@pypi,@github -- uvx fcvm-mcp   # agents can only reach PyPI and GitHub

In a client's JSON configuration:

{"mcpServers": {"fcvm": {"command": "uvx", "args": ["fcvm-mcp"]}}}

The launcher looks for fcvm on PATH, then in ~/.local/bin and /usr/local/bin. Set FCVM_BIN to use another one. Running fcvm mcp directly works the same way and needs no Python package.

Settings are environment variables: FCVM_MCP_NETWORK pins the network policy, FCVM_MCP_TEMPLATE picks the default template, FCVM_MCP_TTL and FCVM_MCP_IDLE_STOP stop forgotten sandboxes, FCVM_MCP_SECRETS and FCVM_MCP_FORWARD hand credentials to sandboxes without the agent seeing them. The agents guide covers the tools and every setting.

Apache License 2.0.