接入你的智能体
从发布者处安装。不是 PlugMCP 端点。
远程 URL 还是本机命令? 两种安装方式的区别
A self-hosted MCP server that gives AI agents structured access to OpenVAS / Greenbone vulnerability scanning — without sending your data anywhere.
OpenVAS has no native interface for AI agents. Most integrations require cloud connectivity or expose GVM credentials to every client. OpenVAS-MCP solves this:
See docs/architecture.md for a full architecture diagram and design details.
Run in your terminal:
claude "install this, make no mistake."
If you prefer to stay in control, follow the manual setup below.
Don't have one? Spin up the bundled Greenbone Community Edition stack:
docker compose -f docker/openvas/compose.yaml up -d
Requirements: Python 3.10+
git clone https://github.com/CyberSecAuto-Labs/OpenVAS-MCP
cd OpenVAS-MCP
python3.11 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
Add to mcpServers in your client config file:
{
"mcpServers": {
"openvas": {
"command": "/path/to/.venv/bin/python", // ← edit this to your venv path
"args": ["-m", "openvas_mcp"],
"env": { "GVM_PASSWORD": "secret" } // ← edit this to your GVM password
}
}
}
Config file locations:
Requirements: Docker
Download the compose files from the latest release and run:
# GVM running locally via Unix socket
MCP_API_KEYS="supersecrettoken:my-agent" GVM_PASSWORD=secret docker compose up
# GVM on a remote host via TCP
MCP_API_KEYS="supersecrettoken:my-agent" GVM_HOST=192.168.1.10 GVM_PASSWORD=secret docker compose up
[!NOTE]
MCP_API_KEYSis a comma-separated list oftoken:namepairs sent as a Bearer token by the MCP client. Multiple clients:"tok1:agent1,tok2:agent2". PassMCP_ALLOW_UNAUTHENTICATED=1instead to skip auth on a trusted network.
Point your MCP client at the server:
{
"mcpServers": {
"openvas": {
"url": "http://your-server:8000/sse", // ← edit this to your server address
"headers": {
"Authorization": "Bearer supersecrettoken" // ← your MCP_API_KEYS token
}
}
}
}
[!WARNING] Plain TCP connections (
GVM_HOSTset,GVM_TLSunset) send GVM credentials unencrypted. UseGVM_TLS=1or a Unix socket for anything beyond local dev.
Greenbone Community Edition + MCP server from source in one go:
# Start the Greenbone stack
docker compose -f docker/openvas/compose.yaml up -d
# Start the MCP server, connected via gvmd socket
GVM_PASSWORD=secret docker compose -f compose.yaml -f compose.override.yaml up --build
[!TIP] See
compose.override.yamlfor how the socket volume is mounted.
See docs/configuration.md for the full reference, including TLS options, policy file, scan limits, and logging.
Example: "Scan 192.168.1.0/24 and show me anything above severity 7" — the agent calls create_target → start_scan → get_scan_status → fetch_scan_results(min_severity=7.0).
Example: "Re-run the weekly scan" — the agent calls list_tasks(filter_string="name~weekly") to resolve the task UUID, then start_task → get_scan_status → fetch_scan_results. Unlike start_scan, this adds a report to the existing task's history instead of creating a duplicate task. GVM applies a default page size to list_tasks; pass rows=-1 to return every task.
Example: "Re-run every single-host task whose last report is over a month old and scored above 5" — a single list_tasks call returns every field those three criteria need, with no follow-up call to resolve host counts:
list_tasks(filter_string="severity>5 and last<-1M rows=-1")
# [{"id": "11111111-1111-1111-1111-111111111111", "name": "weekly-dmz-scan",
# "status": "Done", "progress": "-1",
# "last_report": "22222222-2222-2222-2222-222222222222",
# "last_report_date": "2026-08-14T02:15:00+02:00",
# "severity": 6.5, "report_count": 3, "finished_report_count": 3, "trend": "same",
# "target_id": "33333333-3333-3333-3333-333333333333",
# "target_name": "dmz-web-01", "host_count": 1}, …]
The agent then keeps the rows with host_count == 1 and calls start_task on each. A null severity or host_count means unresolved (no report yet, or the target is gone), never zero.
Three things about GMP filters are easy to get wrong, so list_tasks documents them in its tool description and rejects what GVM would quietly mishandle:
Every release image is:
vuln-scan.yml) and at release time, failing on fixable high severity findings (--only-fixed). CPython CVEs whose only fixes land on a release line the image does not ship are filtered out as unactionable — see CI pipelines.startup-egress workflow traces connect() syscalls via strace on every push and PR, asserting no unexpected outbound connections at startup. Integration tests extend this to live GMP code paths.Verify the image signature before running:
cosign verify \
--certificate-identity-regexp "https://github.com/CyberSecAuto-Labs/OpenVAS-MCP/.github/workflows/release.yml@refs/tags/.*" \
--certificate-oidc-issuer "https://token.actions.githubusercontent.com" \
ghcr.io/cybersecauto-labs/openvas-mcp:<version>
MCP server that bridges AI agents to OpenVAS / Greenbone Vulnerability Management.
工具
暂无工具列表。远程服务器在首次健康检查后会获得工具列表。
常见问题
在本页复制适用于 Cursor、Claude Desktop、Claude Code 或 VS Code 的 stdio 配置。它使用发布者的安装方式,而不是 PlugMCP 代理。
此记录未列出必需的环境变量。
OpenVAS 通过 stdio 在本地运行。
上次检查时间:2026-10-27。 MCP server that bridges AI agents to OpenVAS / Greenbone Vulnerability Management.