Archived: the semantic engine that taught AI agents business context

What it does

Wren Engine sat between your data warehouse and AI agents, translating raw tables into business concepts. You modeled your domain in its Modeling Definition Language (MDL); the engine then handled query planning, governance rules, and MCP-based agent access across 17+ connectors including Snowflake, BigQuery, DuckDB, and Spark.

The interesting bit

The core was a Rust context engine built on Apache DataFusion, wrapped with PyO3 bindings and served through a FastAPI-based ibis-server. That architecture—Rust for planning, Python for execution, MCP for agent plumbing—reflected a real bet that agents need semantic understanding, not just schema access. The README’s comparison table is unusually honest about where catalogs, BI tools, and raw text-to-SQL agents fall short.

Key highlights

• Rust + DataFusion query planner with Python bindings (wren-core, wren-core-py)
• MCP server for Claude Code, Cursor, Cline, and other agent environments
• Read-only mode and governance-aware query planning for safer agent usage
• 17+ data source connectors via the ibis-server module
• Now archived and merged into Canner/WrenAI under core/; new development happens there

Caveats

• Repository is archived (read-only) as of the README’s current state; issues and PRs redirect to WrenAI
• The “engine-architecture” diagram is dense and not obviously explained in prose

Verdict

Worth studying if you’re building semantic layers for LLMs—check the MDL concepts and the a16z thesis they cite. Not a standalone choice anymore; go to WrenAI for active development.