项目的 README,来自 github.com/hugues31/duckpad。合并为一个 markdown 文件:README.md。
See the flow. Trust the result.
Duckpad is a free, open-source visual tool for building data pipelines.
Connect steps on a canvas, write SQL or Python where you need it,
and see the data at every step. Powered by DuckDB.
Duckpad is a workspace for preparing and transforming data. You build a pipeline by connecting steps on a canvas: load files or database tables, join, filter, aggregate, export. Each step is a SQL query or a Python class that you can read and edit, and each step keeps its result. You can open any table along the way instead of guessing what the pipeline did.
It runs on your laptop or on a server your team shares. There is no cloud service, no sign-up and no paid edition.
Use it to:
You need Python 3.11+ and Git.
pip install duckpad
duckpad start
Your browser opens Duckpad: choose the administrator password, and you are in.
On macOS or Windows, run duckpad start --trusted-local instead, for your own
projects on your own machine. The default mode confines pipelines to the files
they are allowed to read, which requires Linux with Landlock; see the
platform requirements.
Then choose New project → Start from an example and Run to explore a finished pipeline, or follow Duckpad in 15 minutes to build your first one from scratch.
Click a cell in a result to see how it was calculated: the rows behind a total, through joins and aggregates, down to the original source cells. Coverage is explicit: when a step cannot be explained, Duckpad tells you.

Guided forms cover the common steps: select and rename columns, clean values, filter, join, aggregate. When you need more, open the code. It is plain DuckDB SQL, with the input columns one click away, and you can preview the query you are typing on real rows.
The default palette also includes union, sorting, deduplication, calculated columns, pivot/unpivot, reproducible sampling, ranking, running totals and Parquet export. Every default has its own icon, and its SQL and form parameter types are available to user-created nodes too.

Work on a branch, compare two versions node by node, check the impact on the data and merge after review. Projects are ordinary Git repositories, so you can also push and pull them.

duckpad run or the automation API.The built-in nodes are a starting point. A node is a file: a SQL query with a few comment directives, or a Python class. Duckpad reads the declarations and turns them into ports, a parameter form and a palette entry. There is nothing to compile or register.
A SQL node declares its inputs, typed parameters, outputs and tests in comments, and optionally its icon and color:
-- @label Revenue breakdown
-- @description Total revenue for each value of the column you choose.
-- @icon chart-bar
-- @color green
-- @input orders
-- @param dimension: column(orders) = "region"
-- @param minimum: number = 100
-- @output result
SELECT {{ dimension }}, SUM(amount) AS revenue
FROM {{ orders }}
WHERE amount >= {{ minimum }}
GROUP BY ALL
ORDER BY revenue DESC;
-- @test revenue_is_positive
SELECT * FROM {{ result }} WHERE revenue <= 0;
A Python node is a class with typed attributes. It receives and returns DuckDB relations, and can use the packages your project declares:
import duckpad as dp
from duckdb import col
# @label Split orders
# @icon split
# @color purple
class Split(dp.Node):
minimum: int = dp.param(default=300, ge=0)
column: dp.Column = dp.param(default="amount", input="orders")
class Outputs(dp.Outputs):
accepted: dp.Relation
rejected: dp.Relation
def run(self, orders: dp.Relation) -> Outputs:
return self.Outputs(
accepted=orders.filter(col(self.column) >= self.minimum),
rejected=orders.filter(col(self.column) < self.minimum),
)
Click SQL or Python at the bottom of the palette, paste the code, and the node is ready to drop into a pipeline. Python nodes are available once an administrator enables Python for the project.

More ways to extend Duckpad:
pyproject.toml; Duckpad locks and installs
them.See the SQL syntax and Python nodes guides for the full contracts.
Duckpad sits between visual data preparation tools and SQL frameworks.
| If you come from | You will recognize | What Duckpad does differently |
|---|---|---|
| Visual data preparation tools such as Alteryx Designer, Dataiku DSS or KNIME | A canvas of connected steps, with a result to inspect at each one | Every step, including the guided ones, is SQL or Python you can read, edit and diff. See the vocabulary map for Dataiku and Alteryx users. |
| SQL frameworks such as dbt or SQLMesh | SQL files in Git, tests and environments | A visual canvas, with a result to inspect at each step and cell-level lineage through SQL. Duckpad computes in DuckDB, not inside your warehouse. |
| Notebooks and scripts | SQL and Python, quick iterations on real data | A dependency graph instead of cell order, a form for each step's parameters, and runs you can review and repeat. |
Duckpad is probably not what you need for:
Duckpad is young and moves fast. Releases are frequent, and a breaking one can require a fresh instance: read the release notes before upgrading.
Duckpad is released under the MIT license and has nothing to sell.
pip install duckpad.If Duckpad is useful to you, star the repository and tell a colleague: that is how other people find it.
| I want to… | Guide |
|---|---|
| Install Duckpad or join a team instance | Installation and accounts |
| Learn the basics with a small example | Duckpad in 15 minutes |
| Upload my data and download a result | Use your own data |
| Navigate, save and run a pipeline | Workspace |
| Write SQL or prepare data with forms | SQL syntax · Data preparation |
| Create my own nodes and reuse them | SQL syntax · Python nodes · Components |
| Read from S3 or a database | Connections |
| Check or explain a result | Results · Data tests · Lineage |
| Review changes and automate a version | Git · Jobs · CLI and automation · External agents and MCP |
| Fix a startup, source or execution problem | Troubleshooting |
The documentation index also covers variables, the assistant and instance administration. Reference material includes project files, architecture, validation, releases and security.
Bug reports, ideas and pull requests are welcome. Open an issue or start with CONTRIBUTING.md.