lcm2m-caddis-mcp

An MCP server that exposes the LCM2M Caddis VM2M API to LLM tools like Claude Desktop, Claude Code, and Cursor. Read-only wrappers over equipment, runs, cycles, telemetry, alarms, and more — served as TOON so LLM context stays cheap.

Requirements

• An LCM2M account (username and password)

• Node.js 25+ or Docker

Related MCP server: KVMFleet MCP Server

Use with an MCP client

The server speaks MCP over stdio and works with any MCP-compatible client — Claude Code, Claude Desktop, Cursor, Windsurf, VS Code (Copilot Chat), Zed, Continue, Cline, Goose, and others. Configure your client to launch npx -y @lcm2m/caddis-mcp with CADDIS_USERNAME and CADDIS_PASSWORD in its env. Concrete examples for the two most common config styles follow.

Claude Code (CLI)

claude mcp add caddis \
  --env CADDIS_USERNAME=you@example.com \
  --env CADDIS_PASSWORD='your-password' \
  -- npx -y @lcm2m/caddis-mcp

Claude Desktop, Cursor, and similar (JSON config)

Most clients accept an mcpServers block in a JSON config file — for example claude_desktop_config.json (Claude Desktop) or .cursor/mcp.json (Cursor):

{
  "mcpServers": {
    "caddis": {
      "command": "npx",
      "args": ["-y", "@lcm2m/caddis-mcp"],
      "env": {
        "CADDIS_USERNAME": "you@example.com",
        "CADDIS_PASSWORD": "your-password"
      }
    }
  }
}

Other clients (Windsurf, VS Code, Zed, Continue, Cline, etc.) use similar formats — consult your client's MCP docs for the exact config path and schema. Restart the client after editing config.

Configuration

Available tools

All tools are read-only and prefixed with caddis_. Each maps 1:1 to a VM2M route; responses are TOON-encoded (a compact, JSON-equivalent format mixing YAML-style nesting with CSV-style tables).

• Company: get_company

• Devices: list_devices, get_device

• Equipment: list_equipment, get_equipment, get_equipment_utilization, get_equipment_schedule, get_equipment_cycles, get_equipment_statuslogs, get_equipment_telemetry, get_equipment_shift_history, list_equipment_excessive_downtimes, get_equipment_excessive_downtime

• Org units / tree: get_org_unit, get_org_unit_schedule, list_org_unit_excessive_downtimes, get_tree

• Alarms: list_alarms

• Tags: list_tags, get_tag, list_tag_groups, get_tag_group

• Runs: list_runs, get_run, get_run_cycles

• Status reasons: list_status_reasons

• Catalog: list_manufacturers, list_models, list_categories

Troubleshooting

• Missing credentials — CADDIS_USERNAME / CADDIS_PASSWORD aren't reaching the child process. With docker run -e VAR, VAR must also be set in the parent shell.

• This user belongs to multiple companies… — set CADDIS_COMPANY_ID to one of the numeric IDs listed in the error.

• 401 Unauthorized — bad creds. Test with:

  curl -X POST https://api.lcm2m.com/vm2m/sessions \
    -H 'Content-Type: application/json' \
    -d '{"username":"you@example.com","password":"...","company_id":1}'

• Persistent 429 — raise CADDIS_MAX_RETRY_WAIT_MS or back off the client's call rate.

• Blank Inspector page in Firefox — use Chrome/Brave/Edge, or npx @modelcontextprotocol/inspector --cli ....

Alternative install

The MCP client examples above use npx. To run from Docker or a local clone instead, build using one of the methods below and swap the command/args in your client config.

Docker

git clone https://github.com/LCM2M/lcm2m-caddis-mcp.git
cd lcm2m-caddis-mcp
docker build --target runtime -t lcm2m-caddis-mcp .

Command string:

docker run -i --rm -e CADDIS_USERNAME -e CADDIS_PASSWORD lcm2m-caddis-mcp

Local Node

git clone https://github.com/LCM2M/lcm2m-caddis-mcp.git
cd lcm2m-caddis-mcp
npm install
npm run build

Command string: node /absolute/path/to/lcm2m-caddis-mcp/dist/index.js

How it works

• Auth: first call hits POST /vm2m/sessions → JWT, cached and proactively refreshed 30s before expiry. 401 triggers a single re-login + retry. The Authorization header carries the raw JWT (no Bearer prefix).

• Rate limiting: backend enforces 20 req/10s per endpoint and 60 req/10s per user. On 429, the client parses Retry-After, applies ±20% jitter, and retries up to CADDIS_MAX_RETRIES (capped by CADDIS_MAX_RETRY_WAIT_MS).

• Response format: TOON — a token-efficient JSON dialect that mixes YAML-style nesting with CSV-style tables (~40% fewer tokens than JSON). The raw body is passed through, fenced as ```toon ... ``` so the model sees an unambiguous format boundary.

• Errors: 4xx responses surface as isError: true tool results so the LLM can see the backend error body and recover; 5xx rethrow.

Development

# Create .env.local with CADDIS_USERNAME and CADDIS_PASSWORD (gitignored).
npm install
npm run dev                 # MCP Inspector web UI + tsx
npm test                    # node --test via tsx
npm run build               # tsc -> dist/
npm run typecheck
npm run lint                # biome check

npm run dev opens the MCP Inspector in a Chromium browser (Firefox has rendering issues). Edit source, click Restart, re-run the tool.

Project layout

src/
  index.ts         # MCP server entry (stdio)
  config.ts        # zod env config
  client.ts        # CaddisApiClient: login, retry, rate-limit
  tools/
    schemas.ts     # shared zod helpers + runTool error wrapper
    index.ts       # tool registration
    wrappers/      # 1:1 VM2M route wrappers
    composite/     # higher-level multi-call tools