Ronald te Brake Develops Local MCP Server for Version-Aware Drupal AI Assistance

The proof-of-concept addresses a persistent issue in AI-assisted development: many tools provide answers that are technically plausible but wrong for the specific Drupal version in use. This is especially problematic when working with version-sensitive APIs or hooks that change across releases. By linking AI tools like Cursor and Claude Desktop to locally embedded API docs, te Brake’s MCP server grounds responses in accurate, version-aligned context.

The system uses Ollama to generate vector embeddings from Drupal’s core *.api.php files, which are then stored in a Redis vector database. When an AI query is made, the server detects the project’s Drupal version using composer.lock, runs a vector search, and returns matched documentation blocks and examples from the correct version. This local-first approach eliminates reliance on incomplete training data or external searches and improves result precision.

The server exposes tools such as drupal_search, drupal_lookup, and drupal_app_info, enabling structured, semantic access to Drupal’s internal APIs. In practice, this ensures, for example, that Drupal 10 users are not wrongly advised to use hooks introduced only in Drupal 11. According to te Brake, the project demonstrates how critical accurate context is for effective AI assistance in real-world development.

The MCP server is fully open source and designed to run offline, making it suitable for privacy-conscious or air-gapped environments. Te Brake suggests this experiment could form the foundation of a more comprehensive Drupal AI starter kit or evolve into a shared cloud-based service. The code is available under ronaldtebrake/surge-mcp-dev on GitHub.