MotherDuck Labs

Experimental projects, prototypes, and explorations from the MotherDuck team.

Heads up: Everything in this repo is experimental. Code here is not production-supported, may change without notice, and is published primarily for inspiration, reference, and community tinkering.

Projects

Each experiment lives in its own directory under projects/. Browse the list below or dive into a folder for project-specific READMEs.

Project	What it is
agentic-sql-claude-edition	Single-setup agentic-SQL system for the DABStep benchmark — a compact skill plus a fetch_context semantic layer that discloses domain knowledge progressively. Hits 419/419 on the cleaned test set.
agentic-sql-mini	Minimal A/B harness for the catalog-context-for-agents experiment — can descriptive column names alone replace prose docs as an agent's information source?
bird-bench	Text-to-SQL evaluation framework using the BIRD Mini-Dev benchmark with MotherDuck as the execution backend.
connections-eval-mini	Evaluate AI models on NYT Connections puzzles. Built for the Seattle Startup Summit evals workshop.
controllog	Controllable logging library for AI/agentic systems — events + balanced postings, JSONL transport, optional MotherDuck upload. Shared dep of agentic-sql-claude-edition, agentic-sql-mini, bird-bench, and connections-eval-mini.
controllog-viz	Static HTML views for controllog datasets (JSONL or MotherDuck via one DuckDB layer) — a per-run review with a chain-of-thought conversation explorer and a cross-run dashboard with a run × question progression/regression matrix.
data-chat-mini	Minimal "chat with your data" Next.js app — read-only agentic SQL over the MotherDuck MCP, inline mviz charts, IndexedDB history + local context layer, and controllog telemetry. Runs a fleet on one read scaling token.
metadata_generator	Generate rich metadata for MotherDuck databases — profile statistics, LLM-generated descriptions, and SQL COMMENT statements. Used as a dep by bird-bench.
motherduck-dive-viewer	Generic BI-style Next.js app — sign in with MotherDuck (OAuth 2.1 + PKCE), browse/search your Dives, and view them rendered inline via a server-side query proxy so no MotherDuck token ever reaches the browser.
postgres-vs-motherduck	Same query, same pg driver, two engines side by side — a Next.js page fires one heavy aggregate at managed Postgres and at MotherDuck at once and charts each as it answers, with server-measured latency. Includes a Python pipeline to load Postgres → MotherDuck. Deploys to Vercel.
react-components	Open-source React components extracted from MotherDuck's website and docs — embed a public Dive, embed a private Dive, and an in-browser MotherDuck SQL editor.

Structure

labs/
├── README.md          # This file
├── LICENSE            # MIT
└── projects/
    └── <project>/     # Self-contained experiment with its own README

Each project directory is self-contained: own README, own dependencies, own runtime. There is no shared build system — projects can use whatever language or stack makes sense.

Contributing

These are experiments — issues and discussions are welcome, but expect a slower cadence than production repos. If you want to add a new experiment:

1. Create a folder under projects/<your-experiment>/.
2. Include a README.md covering: what it does, why it exists, how to run it, and known limitations.
3. Open a PR.

License

MIT — see the LICENSE file for the full text. Individual projects may carry their own license; check the project directory.