skip to content
ryu

Driving a whole voice-AI platform from a single prompt

Building OmniDimension's MCP integration layer: two services, 49 tools, an OAuth 2.1 server, shipped to npm and live in production.

1 min read

I shipped an MCP integration layer for OmniDimension: two services, 49 tools, an OAuth 2.1 server, published to npm and running in production.

Any MCP-compatible AI client can now operate an entire OmniDimension account in plain language:

“Create a voice agent that transfers to a human when asked, attach this PDF as its knowledge base, give it a phone number, and start a bulk campaign to my contact list.”

That prompt can create an agent, add a knowledge base, provision a number, and start a bulk call.

Two services, two entry points#

I built it as two services because developers and end users need different doors in.

  • A local npm package (@omnidim-ai/mcp-server) for developers who want it directly inside their IDE.
  • A hosted OAuth service (omnidim-mcp-cloud) for users who just click “Connect” in Claude or any other MCP connector.

The part I cared about most was keeping the MCP tools in sync with our public APIs. New platform capabilities reach AI clients without a second API surface to maintain by hand.

Developer experience#

A tool layer is only as good as the day you have to debug it, so I spent real time there:

  • proper diagnostics tooling
  • local debug logs
  • privacy-safe telemetry
  • production CI and testing before every release

Try it#

It works in any MCP client. The docs walk through connecting Claude Desktop and controlling real agents:

Originally posted on LinkedIn, with a silent demo of Claude driving real agents.