MCP Orchestrator Routes Buyer Requests

▶ Watch (0:21)

Wilson described a client with many buyers: researchers needing GPUs, robotics labs ordering arms. None want to navigate complex ERPs. PwC built a GPT app that communicates with an MCP orchestrator. The orchestrator classifies each request as direct procurement (robotic arms, GPUs) or indirect procurement (an ergonomic chair). It then routes the request to the correct back-end system. The goal is a simple front-end experience with complex orchestration hidden behind MCP.

Skills and Agents Drive Back-End Integration

▶ Watch (2:29)

The MCP orchestrator uses the agent skills format published by Anthropic. It enumerates skills, each with a name and description. When a buyer requests 1,000 GPUs, the orchestrator classifies the intent as direct procurement. It then invokes a direct procurement agent. That agent makes MCP tool calls to the back-end system. It validates the supplier and raises a purchase requisition. The orchestrator returns the confirmation data and tells the GPT app which UI resource to load in the React application.

MCP Gateways Add Granular Security

▶ Watch (6:56)

One back-end system supports OAuth and direct user identity flow. The other only accepts an API key. Wilson noted MCP gateways help secure these connections. The gateway sits in front of the GPT app. It also provides fine-grained control over which tools inside an MCP server can be invoked. For a utility energy client, only specific tools should run in certain scenarios. The gateway adds that layer of policy enforcement.

Pattern: MCP Invokes Agents That Invoke MCPs

▶ Watch (8:15)

Wilson described the pattern as an MCP orchestrator that invokes agents, which then invoke other MCP servers. The orchestrator enumerates skills from markdown files. It invokes the right skill, which triggers an agent. That agent invokes another skill, which calls an MCP server. The pattern simplifies the experience for buyers across research and robotics divisions. They make a request in a simple app. MCP handles all complex routing and back-end integration.

Notable Quotes

So, regardless of who’s coming in, we just need to simplify the experience and then we do all the complex orchestration on the back end using you know, techniques with MCP and agents to kind of do all that together. Jacob Wilson · ▶ Watch (3:20)

So, it’s kind of a pattern of using an MCP to help orchestrate invoke agents, which then also invoke other MCP servers. Jacob Wilson · ▶ Watch (8:40)

Key Takeaways

  • MCP orchestrators can classify buyer intent and route requests to different back-end systems.
  • The agent skills format from Anthropic lets orchestrators enumerate and invoke skills from markdown files.
  • MCP gateways provide fine-grained security policies for tool invocation within MCP servers.