@cloudtori/opsd-mcp

MCP Server adapter for opsd — bridges AI IDEs to the ODP protocol.

Architecture

AI IDE ←MCP (stdio)→ opsd-mcp ←ODP (WebSocket)→ opsd server (LLM-B)

Quick Setup (Recommended)

Install the CLI globally and auto-configure your AI IDE in one command:

npm install -g @cloudtori/opsd

# Configure server connection
opsd init --server http://your-server:8848 --key ops_sk_YOUR_KEY

# Install for your AI assistant (generates rules + MCP config)
opsd init --ai cursor      # Cursor
opsd init --ai windsurf    # Windsurf
opsd init --ai claude      # Claude Code
opsd init --ai kiro        # Kiro
opsd init --ai trae        # Trae
opsd init --ai roocode     # Roo Code
opsd init --ai all         # All assistants

For MCP IDE mode, the CLI only needs server and api_key. You do not need to set local llm_base_url, llm_api_key, or llm_model in opsd-cli unless you also want standalone CLI local-planner mode.

Manual Setup

Cursor / Windsurf / Trae / Kiro / Roo Code

Add to your IDE's mcp.json:

{
  "mcpServers": {
    "opsd": {
      "command": "npx",
      "args": ["-y", "@cloudtori/opsd-mcp"],
      "env": {
        "OPSD_SERVER": "http://your-server:8848",
        "OPSD_API_KEY": "ops_sk_YOUR_KEY"
      }
    }
  }
}

Claude Code

claude mcp add opsd npx -y @cloudtori/opsd-mcp \
  -e OPSD_SERVER=http://your-server:8848 \
  -e OPSD_API_KEY=ops_sk_YOUR_KEY

MCP Tools

| Tool | Description | |------|-------------| | opsd_analyze | Connect + discover server state. Returns session_token. | | opsd_propose | Generate deployment plan, get LLM-B review. | | opsd_revise | Revise plan based on feedback. | | opsd_execute | Commit and execute the plan. | | opsd_resolve | Handle execution problems (retry/skip/abort). | | opsd_status | Server status (OS, runtimes, ports, disk). | | opsd_apps | List deployed applications. | | opsd_logs | Get operation logs. | | opsd_rollback | Rollback an app. | | opsd_poll | Poll deployment progress. |

How It Works

1. AI IDE calls opsd_analyze → MCP server creates an ODP WebSocket session
2. AI IDE calls opsd_propose → sends plan to LLM-B for review
3. AI reviews feedback, optionally calls opsd_revise
4. AI IDE calls opsd_execute → packs project, uploads, commits via ODP
5. Real-time progress via opsd_poll

Environment Variables

| Variable | Description | |----------|-------------| | OPSD_SERVER | opsd server URL (e.g. http://154.219.98.134:8848) | | OPSD_API_KEY | API key from opsd init |

ODP vs MCP

• MCP = Toolkit (stateless, tool schemas)
• ODP = Protocol (stateful sessions, structured messages)

opsd-mcp uses MCP as the entry point but internally drives everything through ODP.

MCP IDE Mode vs Standalone CLI Mode

• MCP IDE mode: your IDE assistant is LLM-A, opsd-mcp is the bridge, and opsd is LLM-B.
• Standalone CLI mode: opsd-cli itself calls a model API as LLM-A using local llm_* config in ~/.opsd/config.json.

These are separate modes. Configuring one does not automatically configure the other.