Dataiku MCP Server

MCP server for Dataiku DSS REST APIs, focused on flow analysis and reliable day-to-day operations (projects, datasets, recipes, jobs, scenarios, folders, variables, connections, and code environments).

Cursor one-click install includes placeholder environment values. Update DATAIKU_URL, DATAIKU_API_KEY, and optionally DATAIKU_PROJECT_KEY after adding the server.

Related MCP server: SAP Datasphere MCP Server

What You Get

• Deterministic normalized flow maps (project.map) with recipe subtypes and connectivity.

• Summary-first outputs with explicit raw/detail toggles where needed.

• Broad test coverage (unit + live integration + optional destructive integration suite).

• Strong error taxonomy in responses: not_found, forbidden, validation, transient, unknown with retry hints.

Tool Coverage

• project: list, get, metadata, flow, map

• dataset: list, get, schema, preview, metadata, download, create, update, delete

• recipe: list, get, create, update, delete, download

• job: list, get, log, build, buildAndWait, wait, abort

• scenario: list, run, status, get, create, update, delete

• managed_folder: list, get, contents, download, upload, delete_file

• variable: get, set

• connection: infer

• code_env: list, get

Prerequisites

• Node.js 20+

• npm

• Dataiku DSS URL + API key

Quick Start

npm ci
npm run build

Run as a local CLI after build:

node dist/index.js

Use directly from npm (after publish):

npx -y dataiku-mcp

Local Build And Testing

Recommended local workflow from repo root:

# install deps
npm ci

# static checks
npm run check

# unit tests
npm test

# build distribution
npm run build

# run MCP server locally (dev)
npm start

Optional live DSS integration tests:

# requires DATAIKU_URL, DATAIKU_API_KEY, DATAIKU_PROJECT_KEY in .env
npm run test:integration

# includes destructive actions (create/update/delete)
DATAIKU_MCP_DESTRUCTIVE_TESTS=1 npm run test:integration

Repository Layout

• src/: MCP server and tool implementations.

• tests/: unit + integration test suites.

• examples/: demos, fixtures, artifacts, and ad-hoc local scripts.

• bin/: package executable entrypoint.

• dist/: compiled output (generated).

Create a local env file:

cp .env.example .env
# then edit .env

Run directly in dev:

npm start

Example scripts and sample outputs are kept under examples/ to avoid root-level clutter.

Environment Variables

• DATAIKU_URL: DSS base URL

• DATAIKU_API_KEY: DSS API key

• DATAIKU_PROJECT_KEY (optional): default project key

• DATAIKU_REQUEST_TIMEOUT_MS (optional): per-attempt request timeout in milliseconds (default: 30000)

• DATAIKU_RETRY_MAX_ATTEMPTS (optional): max attempts for retry-enabled requests (GET only, default: 4, cap: 10)

• DATAIKU_DEBUG_LATENCY (optional): set to 1/true to include per-tool timing diagnostics in structuredContent.debug.latency (off by default)

MCP Client Setup Guide

Use this server command in clients (npm package):

{
  "command": "npx",
  "args": ["-y", "dataiku-mcp"],
  "env": {
    "DATAIKU_URL": "https://your-dss-instance.app.dataiku.io",
    "DATAIKU_API_KEY": "your_api_key",
    "DATAIKU_PROJECT_KEY": "YOUR_PROJECT_KEY"
  }
}

Windows note: if your MCP client launches commands without a shell, use npx.cmd:

{
  "command": "npx.cmd",
  "args": ["-y", "dataiku-mcp"],
  "env": {
    "DATAIKU_URL": "https://your-dss-instance.app.dataiku.io",
    "DATAIKU_API_KEY": "your_api_key",
    "DATAIKU_PROJECT_KEY": "YOUR_PROJECT_KEY"
  }
}

You can also run TypeScript directly during development:

{
  "command": "npx",
  "args": ["tsx", "/absolute/path/to/Dataiku_MCP/src/index.ts"],
  "env": {
    "DATAIKU_URL": "https://your-dss-instance.app.dataiku.io",
    "DATAIKU_API_KEY": "your_api_key",
    "DATAIKU_PROJECT_KEY": "YOUR_PROJECT_KEY"
  }
}

Claude Desktop

1. Open Claude Desktop -> Settings -> Developer -> Edit Config.

2. Add this under mcpServers in claude_desktop_config.json:

{
  "mcpServers": {
    "dataiku": {
      "command": "npx",
      "args": ["-y", "dataiku-mcp"],
      "env": {
        "DATAIKU_URL": "https://your-dss-instance.app.dataiku.io",
        "DATAIKU_API_KEY": "your_api_key",
        "DATAIKU_PROJECT_KEY": "YOUR_PROJECT_KEY"
      }
    }
  }
}

Cursor

Cursor supports both project-scoped and global MCP config:

• Project: .cursor/mcp.json

• Global: ~/.cursor/mcp.json

Example:

{
  "mcpServers": {
    "dataiku": {
      "command": "npx",
      "args": ["-y", "dataiku-mcp"],
      "env": {
        "DATAIKU_URL": "https://your-dss-instance.app.dataiku.io",
        "DATAIKU_API_KEY": "your_api_key",
        "DATAIKU_PROJECT_KEY": "YOUR_PROJECT_KEY"
      }
    }
  }
}

Cline (VS Code extension)

1. Open Cline -> MCP Servers -> Configure MCP Servers.

2. Add this server block in cline_mcp_settings.json:

{
  "mcpServers": {
    "dataiku": {
      "command": "npx",
      "args": ["-y", "dataiku-mcp"],
      "env": {
        "DATAIKU_URL": "https://your-dss-instance.app.dataiku.io",
        "DATAIKU_API_KEY": "your_api_key",
        "DATAIKU_PROJECT_KEY": "YOUR_PROJECT_KEY"
      }
    }
  }
}

Codex / project-level MCP config

This repo already includes a project-scoped MCP file at .mcp.json. The checked-in .mcp.json uses node node_modules/tsx/dist/cli.mjs src/index.ts for cross-platform startup (including Windows); run npm ci first.

NPM Release Workflow

This repo includes a manual GitHub Actions release workflow:

• Workflow file: .github/workflows/release.yml

• Trigger: Actions -> Release NPM Package -> Run workflow

Inputs:

• bump: patch | minor | major

• version: optional exact version (overrides bump)

• publish: whether to publish to npm

Required repository configuration:

• GitHub variable: NPM_RELEASE_ENABLED=true

• Optional variable: NPM_PUBLISH_ACCESS=public

• Trusted publisher configured on npmjs.com for this package/repo/workflow

The workflow will:

1. Install dependencies, run checks/tests, and build.

2. Bump package version and create git tag.

3. Push commit + tag to main.

4. Publish to npm with GitHub OIDC trusted publishing (if publish=true).

5. Create a GitHub Release with generated notes.

Trusted publishing setup (npm):

1. Open https://www.npmjs.com/package/dataiku-mcp -> Settings -> Trusted Publisher.

2. Choose GitHub Actions.

3. Set:

   • Organization or user: clssck

   • Repository: Dataiku_MCP

   • Workflow filename: release.yml

4. Save.

Official MCP Registry

This repo is configured for MCP Registry publishing:

• Metadata file: server.json

• Workflow: .github/workflows/publish-mcp-registry.yml

• Required package field: mcpName in package.json

Server namespace:

• io.github.clssck/dataiku-mcp

Publish paths:

1. Manual: run Publish to MCP Registry in GitHub Actions.

2. Automatic: run the npm release workflow with publish=true (it triggers MCP Registry publish).

Validation notes:

• server.json.name must match package.json.mcpName.

• server.json.packages[].identifier + version must reference a real npm publish.

Recommended Verification Prompt

After adding the server in a client, run:

• project with { "action": "map", "projectKey": "YOUR_PROJECT_KEY" } (defaults to maxNodes=300, maxEdges=600; override as needed)

You should receive a flow summary in text and normalized nodes, edges, stats, roots, and leaves under structuredContent.map. When truncation limits are applied (default maxNodes=300, maxEdges=600), structuredContent.truncation reports before/after node+edge counts and whether truncation occurred.

Notes

• project.map returns a compact text summary; full normalized graph is in structuredContent.map.

• Arrays in normalized map output are deterministically sorted to reduce diff churn.

• job.wait and job.buildAndWait include structuredContent.normalizedState with one of terminalSuccess | terminalFailure | timeout | nonTerminal while preserving raw DSS state.

• With DATAIKU_DEBUG_LATENCY=1, responses include per-tool and per-API-call latency metrics under structuredContent.debug.latency.

• List-style responses are token-bounded by default; use limit/offset (and action-specific caps like maxNodes, maxEdges, maxKeys, maxPackages) to page or expand results when needed.

• dataset.get and job.get are summary-first by default; pass includeDefinition=true to include full DSS JSON in structuredContent.definition.

Sources

• MCP local server connection docs: https://modelcontextprotocol.io/docs/develop/connect-local-servers

• Cursor MCP docs: https://cursor.com/docs/context/mcp

• Cline MCP docs: https://docs.cline.bot/mcp/configuring-mcp-servers