dltHub AI Workbench

AI-assisted data engineering with dlt. Give your coding agent the skills to build, debug, and explore data pipelines.

Works with Claude Code, Cursor, and Codex.

How to get started

Via Anthropic plugin (cold start)

If you don't have dlt or Python set up yet:

1. Add the marketplace in Claude Code: https://github.com/dlt-hub/dlthub-ai-workbench
2. Install the bootstrap toolkit:

   claude plugin install bootstrap@dlthub-ai-workbench --scope project
   

3. Run /init-workspace — it installs uv, creates a Python venv, installs dlt, and sets up AI agent support with dlt ai init.

Via dlt (existing project)

If you already have a Python project:

uv pip install --upgrade dlt[workspace]==1.23.0a1
dlt ai init

dlt ai init auto-detects your coding agent (Claude Code, Cursor, or Codex) and installs shared rules, secrets handling, and the workspace MCP server.

Then install your first toolkit:

dlt ai toolkit list                           # see what's available
dlt ai toolkit rest-api-pipeline install      # install one

What is workbench

Workbench is a catalog of toolkits that teach AI coding agents how to work with dlt. It is backward compatible with the Anthropic (and Cursor) marketplace and plugin system.

Each toolkit is an ordered group of skills, commands, rules, and MCP servers. A workflow rule ties them together into a guided sequence — the agent knows which skill to use at each step.

graph TB
    subgraph Agent["AI Coding Agent"]
        A["Claude Code / Cursor / Codex"]
    end

    subgraph WB["Workbench — toolkits"]
        direction TB
        T2["<b>rest-api-pipeline</b><br/>Build, debug &amp; validate pipelines"]
        T3["<b>data-exploration</b><br/>Query data &amp; build reports"]
        T1["<b>bootstrap</b><br/>Environment setup"]
        INIT["<b>_init</b><br/>Shared rules, secrets &amp; MCP"]
    end

    subgraph Components["Toolkit anatomy"]
        direction LR
        SK["Skills<br/><i>step-by-step procedures</i>"]
        CMD["Commands<br/><i>slash commands</i>"]
        RU["Rules<br/><i>always-on context</i>"]
        MC["MCP servers<br/><i>data tools</i>"]
    end

    subgraph DLT["dlt runtime"]
        MCP["MCP Server"]
        CLI["dlt CLI"]
        PIPE["Pipelines &amp; Destinations"]
    end

    A -- invokes --> WB
    WB -. made of .-> Components
    A <-. tools .-> MCP
    MCP --> PIPE
    CLI --> PIPE

Loading

Toolkits

Toolkit	Description	Components
rest-api-pipeline	End-to-end REST API ingestion	8 skills, workflow, MCP
data-exploration	Interactive data analysis and reporting	2 skills
bootstrap	Cold-start environment setup	1 command
_init	Shared rules, secrets handling, workspace MCP	installed by dlt ai init

rest-api-pipeline workflow

The workflow guides the agent through a complete pipeline build:

Step	Skill	What it does
0	find-source	Discover a dlt source for your API
1	create-rest-api-pipeline	Scaffold pipeline code and configure credentials
2	debug-pipeline	Run, inspect traces and load packages, fix errors
3	validate-data	Check schema and data, fix types and structures
4	adjust-endpoint	Production-ready: pagination, incremental loading, schema hints
5	new-endpoint	Add more API endpoints to the pipeline
6	view-data	Query and explore loaded data

data-exploration skills (WIP!)

Skill	What it does
explore-data	Query loaded data with the dlt dataset API and ibis
create-marimo-report	Build interactive marimo notebooks with charts and filters

How to use

Option A: dlt ai command line

The dlt ai CLI manages toolkits and agent configuration. It auto-detects your coding agent and installs components in the right format. When you use this option, toolkits become part of your workspace so you can customize and hack them. This follows the same philosophy as our verified sources.

dlt ai init                                    # set up agent support
dlt ai toolkit list                            # list available toolkits
dlt ai toolkit <name> info                     # show toolkit contents
dlt ai toolkit <name> install [--agent] [--overwrite]
dlt ai secrets list                            # show secret file locations
dlt ai secrets view-redacted                   # print secrets with values masked
dlt ai mcp run [--stdio | --sse] [--features ...]
dlt ai mcp install [--agent] [--features ...] [--name]

Agent auto-detection and install paths:

	Claude Code	Cursor	Codex
Skills	.claude/skills/	.cursor/skills/	.agents/skills/
Commands	.claude/commands/	.cursor/commands/	converted to skills
Rules	.claude/rules/	.cursor/rules/	converted to skills
MCP	.mcp.json	.cursor/mcp.json	.codex/config.toml

Option B: Anthropic marketplace and plugins

Workbench toolkits are standard Claude Code plugins. You can browse and install them directly from the Anthropic marketplace in Claude Code — no dlt CLI needed.

1. Add the marketplace: https://github.com/dlt-hub/dlthub-ai-workbench
2. Boostrap dlthub Workspace. Use dlt ai init to get workspace rules.
3. Browse and install toolkits as plugins
4. Skills and commands appear in your agent immediately

This is the easiest path for Claude Code users who want to get started without touching the terminal.

MCP server

Toolkits that need data access use the dlt MCP server — a read-only interface to your pipelines and destinations, installed automatically with each toolkit.

Tool	Feature	Description
list_pipelines	workspace	List all dlt pipelines in the project
list_tables	pipeline	List schemas and tables for a pipeline
get_table_schema	pipeline	Column names, types, and SQL identifiers
get_table_create_sql	pipeline	Generate CREATE TABLE DDL in destination dialect
preview_table	pipeline	First 10 rows as markdown or JSONL
execute_sql_query	pipeline	Run read-only SQL against any destination

The MCP server uses a pluggy-based feature system. The workspace and pipeline features are built into dlt. External packages (like dlt-mcp) can add more features (e.g. search-docs) via plug_mcp hookimpls — see dlt-mcp#30.

Add and maintain Toolkits

See CLAUDE

License

Elastic License 2.0 — use, modify, and distribute freely. Cannot be offered as a hosted/managed service.