Hyperlight — Claude Code MCP server: isolated Python in a micro-VM
Hyperlight runs an agent's Python in a local micro-VM with its own kernel and hardware isolation. Variables and imports persist from one call to the next.
Apache-2.0 ★ 71 updated 2026-10-24
claude mcp add mcp-server-hyperlight -- uvx mcp-server-hyperlightWhat Hyperlight does
The Hyperlight MCP server boots a Unikraft unikernel in a micro-VM through KVM, MSHV, WHP or HVF, depending on your OS. There is no account, no daemon and no API key; code and data stay on the machine. The guest gets no network unless you allow hosts with --net-allow, and no files unless you mount them.
The image ships CPython 3.12 plus numpy, pandas, scipy, scikit-learn, matplotlib, pillow and requests. On first start it pulls roughly 130 MB and stores a warm snapshot around 800 MB; after that the guest comes up from the snapshot in milliseconds. On Linux with KVM, startup is about 40 ms.
The server is a thin layer over the smolagents-hyperlight executor and talks to a Python script inside the guest through its console. Replies are bounded: 64 MiB total, 2 MiB of printed output, and a tool result keeps at most 100,000 characters. A call that outruns --timeout (30 seconds by default) is stopped and the guest replaced from the snapshot. You need hluk v0.18.0 or newer and a hypervisor; without those the server will not start.
What it helps with
- run agent Python without cloud sandboxes
- keep variables and imports between calls
- isolate code in its own kernel with hardware protection
- install extra packages into the guest with
--package - mount local directories through
--hluk-arg
Example prompts for the agent
- Run this code and show the output
- Install sympy in the sandbox and compute the integral
- Reset the sandbox to a clean state
How to install Hyperlight
Platform CLI
claude mcp add mcp-server-hyperlight -- uvx mcp-server-hyperlightuvx runs the command named after the package; if the README names another one, use uvx --from mcp-server-hyperlight <command>
Platform CLI
codex mcp add mcp -- uvx mcp-server-hyperlightuvx runs the command named after the package; if the README names another one, use uvx --from mcp-server-hyperlight <command>
Config file
{
"mcpServers": {
"mcp": {
"command": "uvx",
"args": [
"mcp-server-hyperlight"
]
}
}
}uvx runs the command named after the package; if the README names another one, use uvx --from mcp-server-hyperlight <command>
Config file
{
"servers": {
"mcp": {
"type": "stdio",
"command": "uvx",
"args": [
"mcp-server-hyperlight"
]
}
}
}uvx runs the command named after the package; if the README names another one, use uvx --from mcp-server-hyperlight <command>
Platform CLI
gemini mcp add mcp uvx mcp-server-hyperlightuvx runs the command named after the package; if the README names another one, use uvx --from mcp-server-hyperlight <command>
Step by step, with commands for each platform: How to add an MCP server to Claude Code, Codex, Cursor, VS Code, Gemini CLI and more
Questions about Hyperlight
How do I install Hyperlight in Claude Code?
Install hluk v0.18.0 or newer first: on Linux and macOS run curl -fsSL https://raw.githubusercontent.com/hyperlight-dev/hyperlight-unikraft/main/install.sh | sh, on Windows in PowerShell run irm https://raw.githubusercontent.com/hyperlight-dev/hyperlight-unikraft/main/install.ps1 | iex. Then add the server with claude mcp add mcp-server-hyperlight -- uvx mcp-server-hyperlight.
What does Hyperlight do?
It gives an agent a Python environment inside a local micro-VM with hardware isolation. Variables, imports and functions persist across calls, like in a notebook.
How does Hyperlight compare to Pyodide and cloud sandboxes?
Pyodide runs in Wasm inside a Deno process and is limited to Pyodide packages; cloud sandboxes want an account and an API key. Hyperlight boots full CPython 3.12 with native wheels on your own machine.