What it does WrenAI sits between your data warehouse and the AI agents you already use — Claude Code, LangChain, Pydantic AI, whatever — and feeds them business context that raw schemas can’t provide. Think column definitions, relationship semantics, past query examples, access controls, and (eventually) the corporate wiki knowledge buried in Notion and Slack. The agent asks in English; WrenAI resolves the intent, recalls relevant memory, validates the plan, and returns governed SQL against 20+ sources.

The interesting bit The project is deliberately agent-driven by design. You don’t configure WrenAI through a web UI; you install “skills” into your existing AI coding agent and let it scaffold the project, enrich context, and run queries. The context itself — MDL definitions, instructions, memory — is versionable and Git-friendly, which is rarer than it should be in this space.

Key highlights

• Apache DataFusion-based engine with 22+ data source connectors
• Modeling Definition Language (MDL) for semantic modeling, metrics, and row/column-level access control
• LanceDB-backed memory with hybrid retrieval for similar query recall
• SDKs for LangChain/LangGraph and Pydantic AI; skills install via npx skills add
• Dry-plan validation and structured error hints before execution
• Core engine merged from archived wren-engine repo; previous GenBI app lives on legacy/v1 branch

Caveats

• The /wren-enrich-context skill is marked “Under development” with grill and auto-pilot modes not yet hardened
• Vision paper and some roadmap docs still live in .tmp/roadmap-discuss/ — the README itself notes this needs cleanup
• WASM and Python bindings exist but the developer experience is clearly optimized for LangChain/Pydantic first

Verdict Worth evaluating if you’re already running agentic workflows against a warehouse and tired of prompt-engineering the same business logic into every new LLM session. Skip it if you want a polished BI dashboard for non-technical users — this is infrastructure, not a product.