Skip to main content

@ddse/acm-mcp

@ddse/acm-mcp integrates Model Context Protocol (MCP) servers as first-class ACM tools. It enables deterministic retrieval of external knowledge during planning and execution.

Installation​

pnpm add @ddse/acm-mcp @modelcontextprotocol/client @ddse/acm-sdk

Connect to MCP servers​

import {McpClientManager, McpToolRegistry} from '@ddse/acm-mcp';

const manager = new McpClientManager();
await manager.connect({
  command: 'npx',
  args: ['-y', '@modelcontextprotocol/server-filesystem', '/tmp']
});

const mcpRegistry = new McpToolRegistry(manager);

You can merge MCP tools with local tools:

import {CombinedToolRegistry} from '@ddse/acm-mcp';

const toolRegistry = new CombinedToolRegistry(localTools, mcpRegistry);

Supported servers​

  • @modelcontextprotocol/server-filesystem
  • @modelcontextprotocol/server-github
  • @modelcontextprotocol/server-brave-search
  • @modelcontextprotocol/server-memory

Supply tokens via environment variables (GITHUB_TOKEN, BRAVE_API_KEY, etc.).

Context directives​

The nucleus issues directives such as filesystem:/tmp/report.json. Register matching functions to translate directives into MCP tool invocations and promote results into the context packet.

contextProvider.register(mcpRegistry.lookup('filesystem'), {
  match: directive => directive.startsWith('filesystem:'),
  buildInput: directive => ({ path: directive.slice('filesystem:'.length) })
});

Best practices​

  • Run MCP servers in the same process as demos (via --mcp-server CLI flag) or out-of-process for production workloads.
  • Budget API calls using the policy engine to prevent runaway retrieval.
  • Record MCP responses in the ledger for auditability.

References​