OpenPipeline

From SELECT * to plain English.

OpenPipeline turns every dbt model into a story anyone can read. Every change, from a teammate or an AI agent, goes to the data's owner before it lands.

Website · Live demo (read-only, nothing to install) · Docs

Agents propose, owners approve. AI agents are starting to write SQL. OpenPipeline lets them, safely: an agent reads your project, checks its idea on sample rows, and files a proposal. The owner sees it on the graph, in plain words and business terms, and accepts or rejects it. Nothing touches your dbt files until a person says yes.

It opens a dbt project you already have. It never runs dbt, never connects to your warehouse, and reads and writes plain dbt files, so you can stop using it at any time.

How a change lands

1. Propose. A teammate opens a pull request, or an agent calls propose_change through the MCP server.
2. Understand. OpenPipeline shows what the change does in plain English (or in your own business words), which models and columns it touches downstream, and which business rules and owners are involved.
3. Approve. The owner accepts or rejects it, with comments. Only then is a file written.

Try it in one minute

pip install openpipeline
openpipeline demo

This copies the bundled jaffle_shop example to a temporary folder, files one pending proposal from an imaginary agent, and starts the server at http://127.0.0.1:8484. Open the Proposals tab and press Show on graph to see a review.

What you get

• Model graph. Every model and source as an interactive map. Click one to light up everything it reads from and everything that depends on it.
• Story view. Each model explained step by step in plain words: where the data comes from, how it is joined and filtered, what one row means, and how each column is worked out.
• AI explanations (optional). Bring your own Anthropic key to get a friendlier write-up and suggested descriptions. Nothing is sent unless you click the button.
• Editable docs. Fix descriptions in the UI. They are saved into your own YAML files and the comments in them are kept.
• Sample rows. Run a model on small local sample data in an in-memory DuckDB and see real rows. No warehouse involved.
• Search and column lineage. Press / to find any model, source, or column. Pick a column to see where it comes from and what uses it.
• Business rules. Name the rules buried in your SQL, give each an owner, and link them to the models and columns that carry them.
• PR summaries. A GitHub Action comments on pull requests with what changed in business terms, what is affected downstream, and which owners to loop in.
• Safe changes from AI agents. An MCP server lets agents such as Claude read the project and propose changes. You review each proposal on the graph, comment, and accept or reject it. Nothing is written until a person accepts.

Use it on your own project

openpipeline serve path/to/your/dbt/project
# open http://127.0.0.1:8484

With Docker:

docker run --rm -p 8484:8484 -v "$PWD:/project" ghcr.io/pawanpatil19/openpipeline
docker logs <container>   # prints the write token and a link that opens with it

The container listens on all interfaces, so changes (editing docs, accepting proposals, comments, AI calls) need a token. It makes one up and prints it, or use -e OPENPIPELINE_TOKEN=.... Add -e OPENPIPELINE_READ_ONLY=1 for a server that refuses every write. See Write access and security.

You do not need dbt installed or warehouse credentials. OpenPipeline reads target/manifest.json when it exists (run dbt compile for the most accurate results) and otherwise reads your SQL and YAML files directly.

Optional extras:

pip install 'openpipeline[ai]'        # AI explanations, with your own Anthropic key
pip install 'openpipeline[preview]'   # sample rows, run locally with DuckDB

Connect an AI agent

claude mcp add openpipeline -- openpipeline mcp /path/to/dbt/project

The agent can look at everything, but it can only propose edits. It reads with explain, search, column_lineage, impact_of, and list_rules, checks its work with preview_change and sample_rows, and files changes with propose_change or propose_docs, and suggests glossary terms with propose_glossary_term. See Working with AI agents.

Command line

openpipeline demo                        # the example project plus a proposal to review
openpipeline serve [PROJECT]             # web UI and API on :8484 (--token, --read-only)
openpipeline explain customers [PROJECT] # explain one model (add --ai for an AI write-up)
openpipeline lineage customers lifetime_value [PROJECT]
openpipeline sample customers [PROJECT]  # a few rows, on sample data
openpipeline rules [PROJECT]             # business rules, grouped by owner
openpipeline review BASE HEAD            # Markdown summary of a change, for a PR
openpipeline mcp [PROJECT]               # MCP server for agents

PROJECT defaults to the current folder or $OPENPIPELINE_PROJECT_DIR. Run any command with --help for its options.

Documentation

Guides, the command line, HTTP API, and configuration reference are at https://openpipeline.vercel.app/docs/.

How it is built

Part	Path	Stack
Engine	packages/core	Python, sqlglot, pydantic
App (CLI, API, MCP)	apps/server	FastAPI, Typer, MCP Python SDK
Web UI	apps/web	React, TypeScript, Vite, React Flow
PR summary action	actions/pr-summary	GitHub composite action
Example project	examples/jaffle_shop	dbt

Read docs/architecture.md for how the parts fit.

Status

Early (0.x). The features above work and are covered by tests, including on real public dbt projects, but names and options may still change before 1.0. Bug reports and ideas are welcome in the issue tracker.

Contributing

CONTRIBUTING.md explains how to set up the project and run the checks. In short:

make install   # Python and web dependencies, plus git hooks
make dev       # API with reload on :8484, web dev server on :5173
make lint typecheck test

License

Apache 2.0