@tsa-ai/opc-mcp-server

Local stdio MCP bridge for the OPC Suite. Connects Claude Desktop, Claude Code, Codex CLI, Cursor, Cline and any other Model Context Protocol client to a user's OPC workspace using a personal access token (PAT).

The bridge is a thin pipe — every JSON-RPC frame is forwarded over HTTPS to the OPC instance's hosted MCP endpoint at /api/mcp, which holds the actual tool catalog. Upgrading the toolset on the server immediately benefits every connected agent without users having to update this package.

Setup

1. Sign in to your OPC instance and open Dashboard → Settings → Developer API → AI Connections.
2. Pick the AI tool you want to connect, choose the scopes it needs, and generate a token.
3. Add the snippet below to that tool's MCP config (the in-app wizard prints the exact config for each tool).

{
  "mcpServers": {
    "opc": {
      "command": "npx",
      "args": ["-y", "@tsa-ai/opc-mcp-server"],
      "env": {
        "OPC_BASE_URL": "https://app.your-opc.com",
        "OPC_API_TOKEN": "opc_pat_xxxxxxxxxxxxxxxx",
        "OPC_AGENT_KIND": "claude"
      }
    }
  }
}

Environment variables

| Variable | Required | Default | Description | | ----------------- | -------- | -------- | -------------------------------------------------------- | | OPC_BASE_URL | yes | — | Base URL of the OPC instance (no trailing slash) | | OPC_API_TOKEN | yes | — | opc_pat_… token issued via the AI Connections wizard | | OPC_AGENT_KIND | no | stdio | Used in audit logs to identify the calling agent |

Tools exposed

The catalog lives on the server side (single source of truth) and currently includes:

• Contacts — list_contacts, get_contact, create_contact, update_contact
• Products & inventory — list_products, get_product, list_inventory, adjust_inventory
• Sales — list_sales_orders, create_sales_order, list_quotations, create_quotation
• Accounting — get_profit_and_loss, get_balance_sheet, get_aging_report, get_accounting_summary
• AI Advisor — ask_opc_advisor
• Agent cowork board — create_agent_task, list_agent_tasks, claim_agent_task, append_agent_task_note, complete_agent_task

Each tool is gated by an OAuth-style scope on the user's PAT, so users grant only the permissions a given AI tool needs. The current scope on the token also constrains the result of tools/list, so agents only see tools they can actually call.

Audit & retention

Every tool call is logged on the server with the calling agent, latency, and status (ok / denied / error). Users can review activity in the Usage tab of Developer API settings. Audit rows are retained for 90 days; completed agent tasks for 180 days.

Local development

cd packages/opc-mcp
npm install
npm run build
OPC_BASE_URL=http://localhost:3000 OPC_API_TOKEN=opc_pat_... node dist/index.js

The bridge only forwards traffic — to add or modify tools, edit lib/mcp/registry.ts in the OPC repo, not this package.

License

MIT