Model Context Protocol (MCP)

This feature is currently incubating

The Model Context Protocol (MCP) is an emerging communication protocol for AI-enabled tools to interact with external services.

Services in the Service Library can be exposed to an AI Agent using the Streamable HTTP or Server-Sent Events (SSE) protocols.

The MCP standard deprecated the Server-Sent Events (SSE) protocol and recommended using Streamable HTTP.

MCP server for service

Each service can be exposed individually through MCP. An AI agent can then interact with that service to invoke it and gather the results.

Today, our MCP implementation supports:

  • All decision services.

  • Straight through workflow services.

  • Long-running workflow services and case services that can be resumed at a later time.

  • Workflow services that interact using message resume points.

Add good model descriptions. These are exposed to the AI agent to guide it on how to use the service.

Long-running services support

Workflow/Case services that are long-running can be resumed at a later time. The AI agent can interact with the service to resume the workflow and gather the results.

MCP server exposes following tools for each long-running service:

  • Find service instances that are in progress and can be resumed by tags - find_service_instances

  • Load service instance by ID - get_service_instance

  • Resume service instance by ID and resume point ID - resume_service_instance

  • Get resume point context and schema for a service instance by ID and resume point ID - get_service_instance_resume_context

MCP tool for resuming a service instance validates input to ensure that provided data is in expected format. In case validation fails, error messages are returned and advice is given to load resume context to understand the expected data format.

MCP server for environment

All services of the environment can be exposed through MCP. An AI agent can then interact with each service to invoke it and gather the results.

A popular AI agent implementing MCP is Microsoft CoPilot Studio. They provide a complete documentation on how to deploy services through MCP.

Microsoft CoPilot Studio ask you to split the URL in a swagger YAML definition of the endpoint. You can follow their documentation to split it.

OpenAI, Antoropic and many other also support the MCP protocol.

Different AI agent provider may not have the exact same interaction capabilities and may interact with the service in an inconsistent or incorrect way.