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.

Related MCP server: MCP Auth Server

Features

• Declarative registry — RESOURCES table + _make_list/_make_show/_make_update/_make_delete generators; 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 — *_delete tools 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=True for all fields. limit=N caps row count. all_projects=True for the admin view where supported.

• Multimount — 7 per-domain FastMCP instances served at /<domain>/mcp, each carrying a routing map in its initialize instructions so clients pick the right mount on the first try.

• Kolla log observability — log_targets, log_tail, log_trace read 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:

1. Add a function in src/server.py or src/os_backend.py using openstacksdk.

2. Register it with add(fn, name="...", domain="...", tier="write").

3. For a full CRUD resource, add one dict to RESOURCES and a RESOURCE_DOMAIN mapping entry; _make_list/_make_show/_make_update/_make_delete generate 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.