finds.dev← search

// the find

Canner/WrenAI

★ 15,524 · Python · NOASSERTION · updated Jun 2026

Give AI agents the context to query business data correctly through the open context layer that gives AI agents grounded, governed memory, context, SQL across 20+ data sources, that helps you build agentic GenBI, text-to-sql, dashboards, and agentic analytics.

WrenAI is a semantic layer for AI agents — it sits between your agent and your database, translating natural language questions into governed SQL via a Modeling Definition Language (MDL) that captures business context, relationships, and access rules. The pitch is that your agent shouldn't have to rediscover what 'revenue' means from raw schema every time. It's aimed at teams building text-to-SQL features or agentic analytics pipelines who are tired of hallucinated joins.

The core engine is Rust (Apache DataFusion) with Python bindings and a WASM build — the query planner is not an afterthought, and the sqllogictest suite against TPC-H benchmarks shows they're serious about SQL correctness. MDL as a versionable, Git-diffable artifact is the right call: define business semantics once, share across every agent without vendor lock-in. The dry-plan validation before execution is genuinely useful — catching bad SQL before it hits your warehouse is worth the round trip. Twenty-two connector implementations with a clear factory pattern means you can swap data sources without rewriting your semantic definitions.

The README ships with multiple placeholder comment blocks still in it (`<!-- TODO: vision_paper_en.md is currently at .tmp/... -->`, a hero demo section that's just commented-out instructions) — not a great signal about polish. The 'governed execution' features (audit logs, rate limits, approval workflow) are roadmap items, not shipped; if you're adopting this for anything compliance-adjacent you're building on foundation that doesn't exist yet. The access control story (RLAC/CLAC) appears in the feature list but the 'What's next' section still lists 'full governed execution' as upcoming — evaluate that gap carefully. The agent-driven onboarding UX where your AI agent reads a workflow guide and configures itself is clever in demos but fragile in practice: one misread step puts you in a debugging loop with no deterministic fallback.

View on GitHub → Homepage ↗

// want more like this?

We dig through GitHub every week and send a few repos picked for what you actually care about — each with an honest take like this one.

Get finds in your inbox → Search again →