Skip to main content

MCP Support Overview

You can expose a worker's capabilities to Model Context Protocol (MCP) clients such as Claude and ChatGPT. You declare a tool or a resource, implement one method, and the framework serves it over MCP — you never touch the protocol, transports, or content blocks.

note

MCP support is available from @picsart/pa-pluggable-workers-core 7.6.0.

What you can build​

  • Tools — actions a client can call. A tool orchestrates the tasks you've already built (via flow.executeTask).
  • Resources — read-only content a client can read.

Enabling MCP​

MCP is off by default. Enable it in two steps.

1. Set the flag in src/config/application.yaml:

mcp:
  enabled: true

2. Import the module in your app module, alongside RestModule:

import { CoreModule, RestModule, McpModule } from '@picsart/pa-pluggable-workers-core';

@Module({
  imports: [CoreModule, RestModule, McpModule.forRoot()],
})
export class RestAppModule {}

When the flag is off, the module is inert — /mcp is not mounted and existing workers are unaffected.

Connecting a client​

The worker exposes MCP over stateless Streamable HTTP at POST /mcp. Through the API gateway, clients connect at:

POST /<mcpServerName>/mcp

<mcpServerName> resolves to your worker by the pa-<mcpServerName>-pluggable-worker naming convention. The gateway forwards the caller's identity, so tools run as the authenticated user just like the REST and Kafka APIs.