A CLI agent that turns your data warehouse into institutional memory

What it does Datus is a Python-based data engineering agent that wraps your databases in a continuously updated context layer. It provides a CLI REPL, web chatbot, REST API, and MCP server for turning natural language into SQL. Under the hood, it uses a node-based workflow engine to handle schema linking, query generation, reasoning, and execution, while a LanceDB-powered RAG store tracks metadata, reference queries, semantic models, and external domain knowledge.

The interesting bit The project treats data engineering less as pipeline construction and more as context curation. You can mix and match LLM providers for different workflow nodes—say, a cheap model for schema linking and a stronger one for SQL generation—and package mature domains into “Subagents,” scoped chatbots you hand off to analysts via web, API, or MCP.

Key highlights

• Supports 10+ LLM providers (OpenAI, Claude, Gemini, DeepSeek, Qwen, etc.) with per-node model assignment, plus 11 databases via built-in or pluggable adapters.
• Node-based workflow engine executes steps sequentially, in parallel, or as nested sub-workflows, with agentic nodes for generating semantic models, metrics, and skills.
• LanceDB knowledge base layers schema metadata, reference SQL, Jinja2 templates, metrics, and platform docs to ground generated queries.
• Built-in evaluation against BIRD and Spider 2.0-Snow benchmarks to measure SQL accuracy as context evolves.
• Ships as both an MCP server (for Claude Desktop, Cursor, etc.) and an MCP client, plus a Dashboard Copilot that turns existing BI dashboards into conversational interfaces.

Caveats

• Officially targets Linux and macOS; Windows support is unclear.
• Most database adapters (PostgreSQL, MySQL, Snowflake, ClickHouse, etc.) live in separate Datus-adapters repositories, so setup beyond SQLite or DuckDB may involve extra plumbing.
• The README promises a “continuous learning loop” but does not detail exactly how feedback is ingested or validated.

Verdict Data engineers and analytics teams drowning in ad-hoc SQL requests should look here—especially if you already use MetricFlow or want to expose safe, scoped query agents to business users. If you just need a lightweight SQL CLI without the LLM orchestration, this is overkill.