MCP Dataverse

The most complete MCP server for Microsoft Dataverse.

79 tools · 4 resources · 10 guided workflows · Three auth modes

→ Full Documentation

Why MCP Dataverse?

AI agents hallucinate schema, guess column names, and build broken OData queries. This server gives them real-time access to your Dataverse environment — schema, records, metadata, solutions — through the Model Context Protocol.

• Three auth modes — device code (local), client credentials (CI/CD), managed identity (Azure-hosted)
• Works with any MCP client — VS Code, Claude, Cursor, Windsurf, Gemini, Codex CLI
• Atomic tools — each tool does one thing well; the AI picks the right one
• Structured outputs — every response returns {summary, data, suggestions}
• Guardrails — destructive operations require explicit confirmation
• Encrypted tokens — AES-256-GCM cached credentials, never logged

Install

npx mcp-dataverse install

The interactive wizard configures your environment, registers the server in VS Code, and authenticates your Microsoft account in under 2 minutes.

Requires Node.js 20+. For other clients (Claude, Cursor, Windsurf…) see Multi-Client Setup.

Authentication

Three modes — choose based on where the server runs:

Device code quick start: authentication triggers on the first tool call.

1. Open View → Output → MCP — a sign-in code appears
2. Go to https://microsoft.com/devicelogin, enter the code, sign in with your work account
3. Token is cached encrypted — all future starts are silent

For client credentials and managed identity setup, see Authentication docs.

Capabilities

→ Full Capabilities Reference

HTTP Transport

Run as an HTTP server for multi-client use:

MCP_TRANSPORT=http MCP_HTTP_PORT=3000 MCP_HTTP_SECRET=mysecret node dist/server.js

Connect using VS Code / Copilot with:

{
  "servers": {
    "dataverse": {
      "type": "http",
      "url": "http://localhost:3000/mcp",
      "headers": {
        "Authorization": "Bearer mysecret"
      }
    }
  }
}

Troubleshooting

Performance Tip

MCP Dataverse is designed to be comprehensive, but most AI models work best with fewer tools in context. Deselect the tools you don't need in your client's tool picker (e.g. VS Code Chat panel) to keep the agent focused and responsive.

Roadmap

→ Full Roadmap

License

MIT © Ali Taggaz