OpenStack-MCP
A Model Context Protocol server that provides 123 tools across 7 OpenStack domains (compute, network, storage, etc.) for managing cloud resources via natural language, with stateless per-caller authentication and Kolla log observability.
README
OpenStack-MCP
English | 한국어
A Model Context Protocol (MCP) server for OpenStack — 123 tools across 7 domains, built on openstacksdk, with stateless per-caller header auth, a declarative RESOURCES registry, and Kolla log observability.
Architecture
LLM client (Claude / any MCP host)
│ MCP protocol (stdio or HTTP/SSE)
▼
┌─────────────────────────────────────────────────────┐
│ server.py │
│ RESOURCES table → auto-generated list/show/ │
│ update/delete tools + hand-written specials │
│ │
│ _os_conn(ctx) ──► os_backend.py │
│ (per-caller creds) openstacksdk Connection │
│ (Keystone app credential) │
│ ▼ │
│ OpenStack APIs │
│ Nova · Neutron · Cinder │
│ Glance · Keystone · Octavia │
│ Placement │
│ │
│ ops_backend.py ──► Kolla log files (read-only) │
│ (observability) /var/log/kolla/* │
└─────────────────────────────────────────────────────┘
Per-domain HTTP mounts (stateful sessions for elicitation):
/compute/mcp /network/mcp /lbaas/mcp
/storage/mcp /image/mcp /identity/mcp
/observability/mcp
Each domain is an independent FastMCP instance. A shared process exposes all mounts; OSMCP_DOMAINS and OSMCP_TIERS narrow which tools are active.
Features
- Declarative registry —
RESOURCEStable +_make_list/_make_show/_make_update/_make_deletegenerators; adding a new resource is one dict entry. - Stateless per-caller auth — credentials are read from request headers on every call (HTTP) or from env vars (stdio). The server stores nothing; multiple callers with different credentials share one process safely.
- Structured error envelope — all tool errors surface as
Error executing tool <name>: {"error":{"type","message","http_status?}}. Parse from the first{. - Delete confirmation —
*_deletetools use MCP elicitation to require an explicit human"delete"choice before executing. Irreversible operations cannot be triggered by an LLM alone. - Key-columns / detail — list tools return a compact key-column view by default; pass
detail=Truefor all fields.limit=Ncaps row count.all_projects=Truefor the admin view where supported. - Multimount — 7 per-domain FastMCP instances served at
/<domain>/mcp, each carrying a routing map in itsinitializeinstructions so clients pick the right mount on the first try. - Kolla log observability —
log_targets,log_tail,log_traceread Kolla service log files directly from the host filesystem (mounted read-only), with time-window filtering, regex grep, and request-ID cross-service tracing.
Documentation
- Usage — install, stdio & HTTP modes, container, and configuration reference.
- Tool Reference — all 123 tools by domain.
Quick install:
git clone https://github.com/YeeDochi/OpenStack-MCP.git
cd OpenStack-MCP
python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
Then see Usage to run in stdio or HTTP mode.
Extending: add a new resource
Create tools is intentionally not implemented — it is the primary extension point. To add a create tool or a new resource type:
- Add a function in
src/server.pyorsrc/os_backend.pyusing openstacksdk. - Register it with
add(fn, name="...", domain="...", tier="write"). - For a full CRUD resource, add one dict to
RESOURCESand aRESOURCE_DOMAINmapping entry;_make_list/_make_show/_make_update/_make_deletegenerate the tools automatically.
Any OpenStack service supported by openstacksdk can be wired in this way with a handful of lines.
Running tests
pytest -q
The smoke suite verifies: the tool registry is non-empty and contains the expected OpenStack tools (and excludes non-OpenStack ones), no legacy router module is present, and the Kolla log backend resolves targets/parses request IDs correctly.
License
MIT — see LICENSE.
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。