API Exposure Guide

Model Context Protocol

Native MCP integration for AI agents reasoning over your infrastructure.

Edit this page on GitHub

Model Context Protocol (MCP) Integration

Rescile natively supports the Model Context Protocol (MCP), allowing AI agents (like Claude Desktop) to connect securely and reason over your architecture graph.

The MCP server operates over HTTP Server-Sent Events (SSE) at /mcp and exposes the following tools. Actions defined in your module’s actions/ directory are automatically registered as additional MCP tools — see the Actions Developer Guide for details.

GraphRAG Tools

AI agents can invoke these tools to extract precise data:

  • getResource: Fetch specific properties of a single known node.
  • listResources: Paginated list of resources filtered by type and exact properties.
  • getNeighbors: Retrieve upstream/downside dependencies from a central node.
  • findPath: Find connection paths between two arbitrary resources.
  • localSearch: GraphRAG ego-centric context retrieval leveraging auto-generated community summaries.
  • globalSearch: GraphRAG dataset-wide map-reduce queries.
  • semanticSearch: Sub-string fuzzy matching on resource descriptions and properties.
  • listResourceTypes: Exposes the entire graph schema (node types, property schemas, edge relations).
  • saveAsset (Requires --mcp-write): Allows the AI to modify or insert infrastructure resources directly.

Pre-packaged Prompts

Agents can invoke predefined prompts to streamline interactions:

  • get_resource_details: Analyzes a resource and its direct dependencies.
  • graphrag_local_search / graphrag_global_search: Pre-structures requests for deep graph analysis.