winvm-mcp
An MCP server that gives an LLM full control over a VMware-hosted Windows VM: lifecycle, snapshots, remote execution, file transfer, and kernel debugging.
README
winvm-mcp
An MCP server that gives an LLM full control
over a VMware-hosted Windows VM: lifecycle, snapshots, remote execution
(SSH + vmrun), file transfer, and kernel debugging (cdb.exe / livekd),
all driven from a Linux host over stdio.
Built for offensive and defensive security research on a self-owned lab VM: hunt and prove high-impact vulnerabilities (kernel, drivers, LPE) with full reproducibility through snapshots.
Why
A researcher's loop, identify a target statically -> stand it up in a VM -> set breakpoints -> fire a PoC -> capture the crash or the corrupted state -> write it up, has a lot of tedious context-switching between tools. This MCP collapses that loop into a single conversation: the model drives the VM, the debugger, and the guest shell directly, and consolidates the evidence.
The core architectural choice is that the debugger runs inside the guest over
SSH (via livekd/cdb.exe), so there is no WinDbg-on-Linux dependency and it
works on any host distro.
See docs/ARCHITECTURE.md for the layered design and
docs/THREAT-MODEL.md for the security posture.
Install
The server needs Python ≥ 3.11 on the Linux host, plus vmrun (VMware Workstation/Player). Pick one:
# uv (recommended)
uv tool install winvm-mcp # or, from source: uv pip install -e .
# pipx
pipx install winvm-mcp
# plain pip
pip install winvm-mcp
To run from a checkout instead:
git clone https://github.com/winvm-mcp/winvm-mcp.git
cd winvm-mcp
uv venv --python 3.13 .venv
uv pip install --python .venv/bin/python -e ".[dev]"
Then configure (see below) and point your MCP client at the executable. The
installed entry point is winvm-mcp; it reads its config from the file in
$WINVM_MCP_CONFIG (default: config.toml beside the checkout).
Configure
cp config.example.toml config.toml
$EDITOR config.toml # vmx path, guest credentials, ssh_host, debugger_path
The Windows guest must be set up once. See
docs/SETUP-WINDOWS-VM.md for the full walkthrough
(ISO, network, .vmx tuning, and the guest-bootstrap/ PowerShell scripts that
provision OpenSSH, kernel-debugging, complete dumps, and the livekd launcher).
Sanity-check the server loads your config and registers its tools:
winvm-mcp --version
WINVM_MCP_CONFIG=./config.toml winvm-mcp # then Ctrl-C; it should not error
Register with an MCP client
Generic (stdio)
{
"mcpServers": {
"winvm": {
"command": "winvm-mcp",
"env": { "WINVM_MCP_CONFIG": "/absolute/path/to/config.toml" }
}
}
}
Claude Code / Cursor / Crush / etc.
Use the generic snippet above in your client's MCP config. For a checkout-based
install, point command at the venv Python and args at -m winvm_mcp:
{
"mcpServers": {
"winvm": {
"command": "/path/to/winvm-mcp/.venv/bin/python",
"args": ["-m", "winvm_mcp"],
"env": { "WINVM_MCP_CONFIG": "/path/to/winvm-mcp/config.toml" }
}
}
}
Tools
79 tools across seven groups. Full reference:
docs/TOOL-REFERENCE.md.
| Group | Count | Examples |
|---|---|---|
| VM lifecycle | 8 | vm_state vm_start vm_wait_tools vm_snapshot |
| Snapshots | 4 | snapshot_list snapshot_create snapshot_revert |
| Guest (vmrun) | 10 | guest_exec guest_copy_to guest_capture_screen |
| Guest (SSH) | 2 | ssh_exec_cmd ssh_powershell |
| Kernel debugging | ~35 | kd_connect kd_command kd_memory_read kd_disasm kd_token |
| Analysis workflows | 12 | analyze_dump_overview analyze_live_system |
| Vulnerability research | 6 | vuln_ioctl_dispatch vuln_token_compare vuln_callback_hunt |
Threat model
This server is RCE + kernel-memory-access by design, intended to run on a
researcher's host against a VM they own. Keep the transport local (stdio) and
the VM on an isolated network (NAT / host-only). Never expose the MCP transport
to a network. Full posture and mitigations in
docs/THREAT-MODEL.md.
Development
uv pip install -e ".[dev]" # or: pip install -e ".[dev]"
pytest -q # unit + handshake tests, no VM required
ruff check src tests && ruff format --check src tests
mypy src
python -m build # produce wheel + sdist
Integration tests that need a live VM are marked @pytest.mark.integration and
run only when WINVM_INTEGRATION=1 is set.
License
MIT. No warranties. Use against systems and networks you own or are authorized to test.
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。