One of the first Model Context Protocol (MCP) implementations in the AEC/BIM industry, with an initial prototype built in January 2025, approximately two months after Anthropic publicly released MCP.
The project exposes Autodesk Platform Services (APS) AEC Data Model via a chat-friendly GraphQL gateway, enabling AI assistants like Claude to browse construction projects, query BIM element data, and control a 3D viewer, all through natural language, without requiring users to know APIs or query languages.
Key technical contributions:
- Designed a Skill + MCP layered architecture that routes intent to the correct tool and enforces parameter and identifier rules, making the server more reliable and easier to extend.
- Reduced API round-trips to a single dynamic GraphQL call using RSQL-based element filtering, eliminating the multi-call pattern common in early MCP implementations.
- Added prompt evaluations (Evals) to systematically test tool reliability and prompt correctness, an uncommon practice in AEC tooling at the time.
- Implemented on .NET Aspire with both SSE and stdio transport modes, PKCE OAuth authentication, token caching, and a cross-platform Blazor viewer.
Originated and led development as part of the Autodesk Accelerator programme, where the challenge was brought to the team and collaboratively refined. The work was subsequently recognised in Autodesk’s official APS Developer Blog post “Talk to Your BIM: Exploring AEC Data Model MCP Server with Claude”.
The project was presented during Autodesk DevCon 2026, combining the work of João Martins and Anthony Sertorio
Project opensource in GitHub.
Tools — C#, Python, Claude Code, GitHub, .Net Aspire, Autodesk GraphQL and API
Role — Developer
