VyMCP
MCP server for VyOS routers via VyManager, enabling AI agents to read and manage router configurations with RBAC, audit, and commit-confirm safety.
README
VyMCP
Model Context Protocol (MCP) server for VyOS routers, powered by VyManager.
VyMCP lets AI agents and MCP-aware tools (Claude, IDEs, automation) read and manage VyOS routers through a safe, audited interface. Rather than talking to routers directly, VyMCP wraps the VyManager API — so every action inherits VyManager's per-user authentication, role-based access control, audit logging, and version-aware (VyOS 1.4 / 1.5) configuration handling.
Status: alpha. Read tools cover every VyManager feature; guarded write tools currently cover firewall address groups, with more features being added. The transport is stdio today (hosted HTTP/SSE is planned). Interfaces may still change.
Why a server, not a direct integration
VyMCP deliberately does not connect to VyOS devices itself. It is a thin, stateless translation layer in front of VyManager's REST API. This keeps a single, enforced source of truth for everything that matters in an enterprise environment:
- Per-user identity — a VyMCP request acts as a real VyManager user, never a shared service account.
- RBAC — a token can never exceed its owner's permissions.
- Audit — every configuration change is attributed to the user, token, and instance.
- Commit-confirm safety — writes ride VyManager's existing safety rails.
- Version awareness — VyOS 1.4 vs 1.5 differences are handled by VyManager, not reimplemented here.
MCP client (Claude / IDE / agent)
│ MCP over HTTP+SSE (or stdio)
▼
VyMCP ── stateless translation layer
│ Authorization: Bearer vym_…
│ X-VyOS-Instance-Id: <instance>
▼
VyManager API ── RBAC · audit · commit-confirm · 1.4/1.5 mappers
│
▼
VyOS 1.4 / 1.5
Security model
VyMCP authenticates to VyManager with a personal access token that a user mints in the VyManager UI (Sites → API Tokens). Tokens are least-privilege by design:
- Read-only scope — a token can be restricted so it can never make configuration changes.
- Instance/site scoping — a token can be limited to specific routers or sites (always within the user's own grants).
- Per-call targeting — the target router is selected per request via the instance the token is allowed to reach.
This means you can hand VyMCP a token that, for example, can only read a single production router — and that limit is enforced by VyManager, not by VyMCP.
Tools
Read (always available):
| Tool | Purpose |
|---|---|
list_instances |
The instances this token can reach (returns instance_ids for the other tools). |
list_features |
Every VyOS feature VyMCP can read. |
get_capabilities(feature, instance_id) |
Version-aware capability flags for a feature. |
get_config(feature, instance_id) |
Normalized configuration for a feature. |
Write (only when VYMANAGER_ENABLE_WRITES=true): changes go through a mandatory
propose → apply flow — a propose_* tool validates and returns a plan_id without
touching anything, then apply_change(plan_id, confirm=true) executes that one reviewed
plan. Plans are single-use, time-limited, and bound to their target instance.
| Tool | Purpose |
|---|---|
propose_create_address_group / propose_add_address_group_members / propose_remove_address_group_members / propose_delete_address_group |
Build a firewall address-group change (no effect until applied). |
apply_change(plan_id, confirm) |
Apply a proposed plan, using VyManager's commit-confirm auto-rollback when available. |
confirm_change(instance_id) |
Make a pending commit-confirm change permanent. |
discard_changes(instance_id) |
Revert unsaved changes on an instance. |
get_pending_changes(instance_id) |
Show the commit-confirm rollback window and any pending change. |
Guardrails
- Read-only by default — write tools are not even registered unless
VYMANAGER_ENABLE_WRITES=true, on top of the token's own read-only scope (a read-only token is rejected by VyManager regardless). - No arbitrary config — only curated, typed operations; there is no "set any path" tool.
- Propose → apply —
apply_changerequiresconfirm=trueand a valid plan id, so the model can only apply a change a human has seen. - Never auto-confirms — applied changes ride VyManager's commit-confirm timer and auto-revert unless explicitly confirmed.
Requirements
- A reachable VyManager instance.
- A VyManager API token (Sites → API Tokens; read-only recommended to start).
- One or more VyOS 1.4 / 1.5 routers registered in VyManager.
- Python 3.10+.
Getting started
Install:
pip install -e . # or: pip install -e ".[dev]" for tests
Configure via environment variables (see .env.example):
export VYMANAGER_BASE_URL="https://vymanager.example.com"
export VYMANAGER_API_TOKEN="vym_…" # read-only recommended
# export VYMANAGER_ENABLE_WRITES=true # opt in to write tools
Register with an MCP client (stdio). Example client config:
{
"mcpServers": {
"vymcp": {
"command": "vymcp",
"env": {
"VYMANAGER_BASE_URL": "https://vymanager.example.com",
"VYMANAGER_API_TOKEN": "vym_…"
}
}
}
}
Development
pip install -e ".[dev]"
pytest # deterministic; no live device needed
ruff check .
mypy
Roadmap
- [x] stdio server + token auth to VyManager
- [x] Read tools across all features
- [x] Guarded write tools (propose → apply → commit-confirm)
- [ ] More curated write features
- [ ] Hosted remote transport (HTTP/SSE) with per-user auth
- [ ] Packaging and deployment docs
Related projects
- VyManager — the web UI and API that VyMCP builds on.
License
See LICENSE for details.
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。
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。
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。
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。
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。
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。
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。
mcp-server-qdrant
这个仓库展示了如何为向量搜索引擎 Qdrant 创建一个 MCP (Managed Control Plane) 服务器的示例。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。