The latest features, improvements, and fixes shipped in Bauplan.
SDK 0.3 brings typed data contracts to every pipeline. Each model declares the tables it reads and writes, column by column, so mismatches are caught before any data changes.
Typed schemas: declare each column's type and whether it can be null, directly in the model. Nullability no longer needs a separate expectation.
Semantic annotations: describe what each column means in plain language. Descriptions are stored in the Iceberg tables and versioned with the data.
Column lineage: declare which upstream column a field comes from, so you can trace how data moves through a pipeline.
We published our methodology for optimizing agent skills. It treats a skill's text as a hyperparameter and optimizes it. First it generates a synthetic dataset and then uses GEPA as the skill optimizer. The framework is domain agnostic and not tied to any Bauplan-specific construct.
Read the paper and browse the code.
An alpha release adds semantic annotations that describe the inputs and outputs of your models, giving Bauplan, and the agents working on it, a richer understanding of how data flows through a pipeline.
Job and run logs are now written to your own cloud bucket and retrieved through pre-signed URLs. This removes the previous 24-hour access limit.
In benchmarks against Snowflake and Databricks - adapting Columbia's BranchBench to OLAP - Bauplan creates branches ~100x and ~250x faster, respectively. Our near-instant branch creation lets agents iterate on production data safely and efficiently, running many branches in parallel and looping through them with minimal branch-creation latency.
Read the benchmark post.
A preview integration lets you Cloudflare R2, so you can import Parquet files stored in R2 as zero-copy external tables and run Bauplan pipelines on them.
See the R2 integration docs.
Bauplan's MCP server now supports a standard OAuth flow and works out of the box with Claude Desktop. Connect once and let agents query catalogs, run pipelines, and inspect data. Read about Bauplan's MCP server.
The guide explains how to handle merge conflicts between branches and lays out best practices for resolving them.
Bauplan skills are now packaged as a Claude Code plugin. Add the marketplace, install the plugin, and Claude Code will auto-update a as the skills evolve. Bauplan skills repo.
A new bauplan init command scaffolds a ready-to-run project: the project config, a sample pipeline with a model and an expectation test, and its Python dependencies. See the quick start guide.
Stream data from any Estuary source into Bauplan Iceberg tables. The Bauplan materialization connector merges row-level inserts, updates, and deletes in near real-time. See the Estuary integration docs.
Bauplan now works with Google Cloud in two directions: sync a GCS bucket into your lakehouse with AWS DataSync to run pipelines on GCS-landed data, and register Bauplan tables in BigQuery to query them with no data copy (read-only).
See the GCS and BigQuery integration docs.