sapcc

SAP Converged Cloud Agent Toolkit

Help AI coding agents operate SAP Converged Cloud infrastructure.

The Agent Toolkit gives AI agents the skills, knowledge, and guardrails to work with SAP CC services effectively. It works with Claude Code, Codex, and any agent supporting the skills format.

Prerequisites

Requirement	Purpose
openstack-mcp-server	Runtime providing authenticated API tools (required)
SAP CC credentials	Application credential or username+password for your region
OS keychain	macOS Keychain, Linux pass, or secret-tool for secret storage
Claude Code ≥ 1.0 or Codex	Agent runtime supporting plugin format

MCP Server Setup

# Install the MCP server binary
go install github.com/notque/openstack-mcp-server@latest

# Or download pre-built binary from releases
# https://github.com/notque/openstack-mcp-server/releases

Credential Setup

After installing the plugin, run the credential-setup skill:

/sapcc:credential-setup

This guides you through storing credentials in your OS keychain. Supports:

• Application credentials (recommended) — scoped, revocable
• Username + password — via keychain-backed retrieval commands

Environment variables used by the MCP server:

Variable	Description
OS_AUTH_URL	Keystone endpoint (e.g., https://identity-3.<region>.cloud.example.com/v3)
OS_USERNAME	SAP CC username (if using password auth)
OS_PW_CMD	Command to retrieve password from keychain
OS_APPLICATION_CREDENTIAL_ID	App credential ID (if using app creds)
OS_APPCRED_SECRET_CMD	Command to retrieve app credential secret
OS_PROJECT_NAME	Default project scope
OS_DOMAIN_NAME	Domain (e.g., my-domain)
MCP_READ_ONLY	true (default) = read-only tools; false = enable write tools
MCP_ADMIN_TOOLS	true = enable admin tools (requires cloud_admin role)

Quick Start

Claude Code

/plugin marketplace add notque/openstack-agent-toolkit
/plugin install sapcc@openstack-agent-toolkit

Manual Installation

Copy skills to your agent's skills location:

Agent	Path
Claude Code	~/.claude/skills/ or .claude/skills/
Codex	~/.codex/skills/ or .agents/skills/

Verify Installation

Quick smoke test to confirm everything works:

# 1. Check plugin validates
python3 tools/validate.py --plugin sapcc

# 2. Confirm MCP server starts (should show tool list)
openstack-mcp-server --list-tools

# 3. Test read-only API call (requires valid credentials)
# In your agent, ask: "What project am I authenticated to?"
# Expected: agent calls keystone_token_info and reports project/domain

Architecture

┌─────────────────────────────────────────────────────────────────┐
│  AI Agent (Claude Code, Codex, etc.)                             │
├─────────────────────────────────────────────────────────────────┤
│  Agent Toolkit (this repo)          │  MCP Server                │
│  • Skills (domain workflows)        │  • Tool implementations    │
│  • Knowledge (SAP CC context)       │  • Auth management          │
│  • Rules (guardrails)               │  • API proxying             │
│  • Gotchas (mistake prevention)     │  • Response sanitization    │
└─────────────────────────────────────────────────────────────────┘

MCP Server (openstack-mcp-server) = runtime providing 119 typed tools across 12 services, gated by environment variables Agent Toolkit (this repo) = intelligence layer teaching agents when and how to use those tools

Three-Tier Tool Model

Tools are exposed progressively based on environment configuration:

Tier	Env Variable	Tools	Use Case
Read (default)	MCP_READ_ONLY=true	91 tools	Safe exploration, monitoring, investigation
Write	MCP_READ_ONLY=false	+16 tools	Resource creation, modification, deletion
Admin	MCP_ADMIN_TOOLS=true	+12 tools	Cloud admin operations (hypervisors, agents, chassis)

All write tools enforce a confirmation pattern — they return a preview unless confirmed=true is passed. The destructive-action-gate hook additionally blocks deletes and power-state changes until the user explicitly approves.

Task Routing

When you ask the agent a question, it auto-selects the appropriate skill: