1.2kGet started
Datus MCP Server

Datus MCP — Model Context Protocol Server for Your Warehouse

Expose the agent over MCP so Claude Desktop, Cursor and Windsurf can query, audit and reason about your warehouse with shared context.

Why MCP for a Data Agent?

Every serious AI client — Claude Desktop, Cursor, Windsurf, Cline — speaks Model Context Protocol. Datus-MCP turns your warehouse into a native tool source those clients can call directly.

Data Tools, Not Data Copies

Instead of shipping a snapshot of your schema to the model, Datus MCP exposes tools the host calls on demand — the warehouse stays the source of truth, no stale metadata.

Any MCP Host, One Server

Claude Desktop, Cursor, Windsurf, Cline — every serious AI client speaks MCP. Datus runs once and the same tools show up in whichever host your team prefers.

Governed by the Same Policy

Every tool call passes through the same context, semantic layer and audit log as the CLI. What ships to your host is what your data platform team approved.

MCP Use Cases Across AI Clients

Four host integrations — register Datus once, and the same context tools show up wherever your team already codes.

Ask Claude Desktop About Your Warehouse

Register Datus once. Claude answers “which orders shipped late last week?” by calling query_sql against your Snowflake, with lineage in every reply — the same governed surface API callers hit, on top of the semantic layer data engineers own.

Wire Cursor Into Your Semantic Layer

Editing dbt models? Cursor calls explain_metric to surface the definition, owner and lineage — no leaving the editor to grep for docs. It's the same authoring tool you drive from the CLI, reused by the engineers who author these models.

Reproducible Analysis in Windsurf

Windsurf drafts SQL by calling draft_sql. You review before it runs; every accepted query lands in the workspace transcript with full lineage — the same anomaly investigation loop analysts follow, now inside the IDE, or from the CLI when the pipeline lives in a shell.

Bring Your Own MCP-Compatible Host

Cline, Continue, or a home-grown MCP client — Datus registers once and the same tools show up wherever the team prefers to work, no extra plumbing. Especially useful when self-hosting the open-source stack and pointing your own host at it.

mcp.jsonyaml
{
  "mcpServers": {
    "datus": {
      "command": "datus",
      "args": ["mcp", "serve"],
      "env": {
        "DATUS_WORKSPACE": "growth",
        "DATUS_DATASOURCE": "warehouse"
      }
    }
  }
}

Any MCP Host, One Config

The Datus MCP server ships as a single binary. Register it in your client's config once and the same tool set appears in Claude Desktop, Cursor, Windsurf and Cline — no per-host adapter needed.

MCP Primitives: Tools, Resources, Prompts

The three building blocks of the Model Context Protocol — Datus implements all of them on top of your stack.

TOOL

Callable tools

Actions the host invokes with structured arguments — query_sql, draft_sql, explain_metric, list_tables.

RES

Resources

Read-only references — schemas, metric definitions, past sessions — the host can attach to a conversation.

PROMPT

Prompts

Reusable prompt templates the host can invoke by name — "weekly report", "cohort analysis", "on-call check".

MCP Toolbox: Curated Data Tools for the AI Host

Four opinionated tools that turn a generic AI client into a data-aware assistant.

query_sql

Runs vetted SQL through the context and safety layer. Returns rows plus the lineage used to build them.

list_tables

Browse catalog tables and columns. The host attaches results as MCP resources, ready to reference.

explain_metric

Semantic definition, lineage and owner for a named metric — grounded in the workspace's Context.

draft_sql

Drafts SQL for review before it runs. The host displays diffs; the user approves before execution.

MCP Setup: Three Steps to a Live Tool

Add the server to your host's config, restart, and start invoking Datus tools.

  1. 01

    Install the agent

    Grab Datus from PyPI or Homebrew — the same binary powers the CLI, API and MCP server.

  2. 02

    Register in your client

    Add Datus to your MCP host's config file (Claude Desktop, Cursor, Windsurf, Cline — all use the same shape).

  3. 03

    Restart and pick a tool

    Restart your client, open the tool picker, and start invoking query_sql, explain_metric and the rest.

Pick the Interface That Fits Your Team

Four surfaces. One agent. Pick the one that fits your team.

CLI

Explore data, build context, and ship SQL from the terminal.

Web Chatbot

Chat with subagents from a browser — zero install.

API Server

Consume data services via REST — language agnostic.

You are here

MCP Server

Plug into Claude Desktop, Cursor, and any MCP client.

FAQ

Frequently asked questions

Which MCP clients does Datus work with?

Any MCP-compatible client — Claude Desktop, Cursor, Cline, Continue, and custom clients built on the MCP SDK. Datus supports both stdio and HTTP transports.

What's the advantage over giving Claude raw warehouse credentials?

The Datus MCP server wraps the warehouse with the same context engine, governance and approved-answer cache used by the CLI — so the host AI sees governed, scoped tools instead of an open SQL connection.

Can I scope the server to specific datasources or tools?

Yes. Configure exposed tools, allowed datasources, and read-only mode per server instance. Run separate `datus-mcp` processes for different teams or audiences.

Does the MCP server share context with the CLI and chatbot?

Yes — all four surfaces (CLI, chatbot, API, MCP) read from and write back to the same evolvable context store, so improvements made in one show up everywhere.

Plug Datus Into Any MCP Client

One server exposes your warehouse, semantic layer and catalog as MCP tools — usable from Claude, Cursor, or any agent that speaks the protocol.