Datadef

AIからデータアーキテクチャ図を生成・編集します。

ドキュメント

Datadef MCP

Data-architecture diagrams your AI generates, edits, and exports — from Claude, Cursor, VS Code, Gemini CLI, or any MCP client. Typed tables with columns, pipelines, column-level lineage, 2,000+ real tool icons. Exported PNGs come back inline in the chat.

This repo is the public home of the remote MCP server at https://datadef.io/mcp (Streamable HTTP). There is nothing to install from here — it exists for directories, install links, and issues.

What it draws — live

The diagram below is not a screenshot. It is a public Datadef diagram embedded by URL — when the diagram is edited, this image updates with it. That is the loop this server exists for: your agent draws and maintains the diagram, and every doc that embeds it stays current.

Customer 360 Platform — Identity Resolution and Activation

![Customer 360 Platform](https://datadef.io/api/embed/customer-360-platform)

Connect

Two ways in — both included in the 7-day free trial:

OAuth (no key): add https://datadef.io/mcp as a connector in Claude or ChatGPT and sign in when prompted. Claude Code works keyless too: claude mcp add --transport http datadef https://datadef.io/mcp.
API key (headless): create one at datadef.io/settings/mcp and send it as a Bearer header — right for CI and scripts.

Claude Code

claude mcp add --transport http datadef https://datadef.io/mcp \
  --header "Authorization: Bearer dd_live_YOUR_KEY"

Cursor / Claude Desktop (via the mcp-remote bridge)

{
  "mcpServers": {
    "datadef": {
      "command": "npx",
      "args": [
        "-y", "mcp-remote", "https://datadef.io/mcp",
        "--header", "Authorization: Bearer dd_live_YOUR_KEY"
      ]
    }
  }
}

VS Code (.vscode/mcp.json)

{
  "servers": {
    "datadef": {
      "type": "http",
      "url": "https://datadef.io/mcp",
      "headers": { "Authorization": "Bearer ${input:datadef-key}" }
    }
  }
}

Gemini CLI (~/.gemini/settings.json) — note httpUrl, not url; plain url is legacy SSE there and fails silently:

{
  "mcpServers": {
    "datadef": {
      "httpUrl": "https://datadef.io/mcp",
      "headers": { "Authorization": "Bearer $DATADEF_API_KEY" }
    }
  }
}

One-click install buttons live at datadef.io/settings/mcp.

Tools

Three layers. Outcome-level — describe an intent, Datadef's pipeline carries it out: create_diagram, list_diagrams, get_diagram, edit_diagram, export_diagram, get_design_guide. Repository sync — repo_status and repo_refresh inspect and re-run the sync that keeps a diagram and its architecture.md regenerated from a connected GitHub/GitLab/Azure DevOps branch or tag. Terraform repositories get a dedicated pipeline: every .tf file parsed (no init, no state, no cloud credentials), modules drawn as zones, per-environment counts kept honest. Atomic — 25 canvas_* tools your model drives directly (add/update/remove nodes, connect edges, set columns, add lineage, group, align, layout, validate), so the agent that already knows your repo can draw what it finds. The datadef_design_guide prompt teaches any model the design standard before it draws.

Anonymous initialize and tools/list are open — point any MCP inspector at the endpoint to browse the surface before creating a key.

Why a diagram tool wants an agent

The agent that just changed your dbt project still has the whole change in context. Telling it "update the architecture diagram too" costs one sentence — and an embedded diagram (![arch](https://datadef.io/api/embed/your-diagram)) updates everywhere the canvas does. Docs stop rotting.

Run as a container

For clients or checkers that want a runnable image, the included Dockerfile bridges stdio to the hosted server:

docker build -t datadef-mcp .
docker run -i -e DATADEF_API_KEY=dd_live_YOUR_KEY datadef-mcp