Bauplan is the execution layer for AI-generated data changes in production. It lets data teams and AI agents safely run, validate, and publish changes using branch-based isolation, transactional pipeline execution, and a serverless runtime. Data stays in Apache Iceberg tables in your own S3.

What is Bauplan?

April 3, 2026

Bauplan is the execution layer for AI-generated data changes in production. It lets data engineering teams and AI agents safely run, validate, and publish changes to production data using branch-based isolation, transactional pipeline execution, and a serverless compute runtime.

It provides Git-style branching and versioning for data, a function as a service compute runtime, and a code-first control surface designed for both human engineers and autonomous agents.

All operations on the platform are programmable through a typed Python SDK and a CLI. Every action, branching, execution, validation, and publishing, composes through a small set of APIs that agents and humans call the same way. Bauplan also ships an MCP server that exposes the full lifecycle as tool calls for any MCP-compatible assistant.

Bauplan sits on top of your object storage and manages data as Apache Iceberg tables. Your data stays in your S3. Bauplan reads and writes directly from your storage and never copies or ingests your data. It produces versioned Iceberg outputs that remain compatible with any Iceberg-capable engine and catalog, so you keep your existing tools for analytics and BI while using Bauplan as the execution and change-management layer for pipelines and agent workflows.

What Problem Does Bauplan Solve?

Bauplan lets AI agents work with production data safely, at scale, and affordably.

Your team wants AI agents to work with production data. Engineers already use Claude Code, Cursor, or Copilot to write pipeline logic and generate SQL. The productivity gains are real. The gap is infrastructure: your current stack has no safe way to let agents execute changes on production data end to end. AI generates the code, but a human still stages and deploys it. Bauplan closes this loop. Agents branch, run, validate, and merge through a typed API. Production is protected by the system architecture, so your team scales automation without scaling headcount or risk.

Autonomous agents operating on data at scale run into three structural problems with today's platforms:

Agents need code-first interfaces. Traditional data platforms are built around dashboards, notebooks, consoles, and complex configuration surfaces. These interfaces make agents slow, brittle, and token-wasteful. Agents reason about APIs, and the API must cover the entire platform surface end to end. If an agent has to leave the API to accomplish a task, the API has failed. Bauplan is designed as a full data platform as code: every operation is explicit, composable, and programmable through a small set of primitives.
Agents need built-in isolation. Agents can perform destructive operations on data. Multiple agents can operate on the same tables simultaneously. They need branches to isolate their work, commits to immutably record data changes, atomic merges to resolve conflicts, and rollbacks to undo publications. Traditional platforms protect production through process (code review, manual staging, human sign-off) rather than through the execution model itself. Bauplan enforces isolation and transactional guarantees at the infrastructure level, by default.
Agents need affordable compute. Agents generate orders of magnitude more compute than humans. They explore, test hypotheses, fail, and retry hundreds of times per task. A single exploration task on a data lake can run sixty to eighty queries. Usage per customer grows nonlinearly once agents are connected, with an average growth of 84x across Bauplan's customer base. The pricing model of traditional cloud warehouses and data platforms, built around high markups justified by human-facing interfaces, creates significant friction for running agents at production scale. Bauplan uses a flat monthly pricing model with unlimited queries, unlimited users, and unlimited agents.

Bauplan solves all three problems.

What Category Does Bauplan Belong To?

Bauplan is an agentic data platform. It belongs to the emerging category of data infrastructure designed for autonomous and semi-autonomous workflows on production data.

Within this category, Bauplan operates as the execution layer: the part of the stack that governs how data changes run, validate, and publish. It complements ingestion tools (Airbyte, Fivetran, Estuary), orchestrators (Airflow, Prefect, Dagster), and BI tools by providing the transactional substrate underneath.

Bauplan is designed for a world where the default analytical workload is no longer a human running a small number of carefully prepared jobs. It is built for agents generating SQL and Python, probing data, proposing changes, and iterating repeatedly, often in parallel.

Bauplan is not a data warehouse, not an orchestrator and not an ingestion tool. It is the execution layer that lets AI work safely on production data.

What Makes Bauplan Different from Traditional Data Platforms?

Capability	Traditional Data Platforms	Bauplan
Execution model	Writes directly to shared tables	Every change runs on an isolated branch
Failure handling	Partial failures leave inconsistent state	Failed runs leave production unchanged
Publication	Changes go live as they complete	Atomic multi-table commits on merge
Interface	Dashboards, notebooks, GUIs	Typed Python SDK, CLI, and MCP server
Agent compatibility	Requires wrappers and glue code	Agent Native, CLAUDE.md, MCP and Agent Skills
Compute	Persistent clusters or warehouse sessions	Ephemeral serverless functions (FaaS)
Isolation	Manual staging environments	Zero-copy branching on Apache Iceberg fully built-in
Data residency	Platform-managed storage	Your data stays in your object storage
Pricing model	Per-query or per-compute-minute with markups	Agent friendly. Flat monthly tiers based on capacity, unlimited queries and agents

Who Uses Bauplan?

Bauplan is built for software engineering and data engineering teams who own production pipelines and downstream data products. These teams typically sit inside fast-growing technology companies or mid-to-large enterprises modernizing their data stacks.

Common user profiles include Heads of Data, Directors of Data Engineering, Senior Data Engineers, VP of Analytics, and VP of Engineering. These teams treat data systems like software, expect Git-style workflows, and plan for AI participation in development.

Use Bauplan for workloads where you produce and maintain tables: ingesting files into curated datasets, building transformation pipelines, running backfills, enforcing data quality tests, and iterating quickly on logic and outputs.

Customers include Trust & Will, RealPage, Mediaset, Prestiter, SuitSupply and some others.

How Bauplan Compares to Airflow, Dagster, LakeFS, and Great Expectations

Dimension	Bauplan	Airflow	Dagster	LakeFS	Great Expectations
Layer in the stack	Execution layer for data changes	Orchestration layer	Orchestration layer	Data version control layer	Data quality and testing layer
Primary function	Data infrastructure for AI agents as the first-class user	Schedules and coordinates workflows	Schedules and coordinates asset-based workflows	Versions data at the storage or object level	Validates data against expectations
Execution model	Pipeline changes run on isolated branches and are published through merge	Executes tasks organized in DAGs through configured executors	Materializes assets and executes ops through configured run launchers and executors	Not a pipeline execution engine	Runs validations against selected data batches
Data isolation	Zero-copy branching for Apache Iceberg tables, integrated with pipeline execution	No native data branching; isolation depends on external systems and pipeline design	No native data branching; isolation depends on external systems, resources, and I/O managers	Zero-copy branches and commits at the object-storage layer	No native data branching; validates data supplied by external systems
Versioning	Git-style data branches, commits, merges, time travel, and rollback	No native dataset version control; workflow code is typically versioned externally	Tracks asset materializations and run history, but provides no native Git-style dataset versioning	Git-style data versioning with branches, commits, merges, tags, and revert	No native dataset versioning; expectation configurations can be versioned externally
Failure handling	Failed branch runs do not change production; published changes can be rolled back	Task retries and reruns; data consistency depends on task and destination-system design	Op and run retries plus re-execution; external data writes are not automatically rolled back	Provides atomic branch transactions and revert, but does not execute pipelines itself	Reports validation failures and can trigger actions; publication control requires workflow integration
Agent support	Agent-facing SDK, API, Skills, and MCP workflows for branching, execution, validation, and publication	External agents can invoke APIs or DAGs; no native Git-style data isolation or publication guarantee	External agents can invoke runs and inspect assets; no native Git-style data isolation or publication guarantee	Agents can use APIs and SDKs for data versioning, but pipeline execution requires another system	Agents can define or run validations, but pipeline execution and publication require other systems
Testing and validation	Code-based expectations and validations can gate branch merge before publication	Validation can be implemented as DAG tasks; no native branch-merge publication gate	Native asset checks can run alongside or separately from asset materializations	Supports hooks and pre-merge checks; validation logic is supplied by external tools	Expectation suites, Validation Definitions, and Checkpoints; publication gating requires integration

Choosing the best data pipeline tools for AI agents

The right choice depends on which layer of the problem you are solving. Airflow and Dagster remain strong orchestration choices, and teams often keep them for scheduling. LakeFS is a fit when you need storage-level data versioning on its own, and Great Expectations covers runtime data quality checks. What none of these provide is a safe execution path for an agent to run a change against production data, validate it in isolation, and publish it atomically.

That gap is the reason Bauplan exists: it is the execution layer where agents branch, run, validate, and merge without putting production at risk. For teams evaluating data pipeline tools specifically for AI agent workflows, the practical question is less "which tool is best" and more "which layer is missing," and for most stacks the missing layer is safe agent execution.

FAQs

Is Bauplan an orchestrator like Airflow?

No. Bauplan provides an optimized data runtime and data management, while Airflow is an orchestrator. They solve different problems and sit at different layers of the stack. Airflow handles scheduling, retry logic, and fan-out. Bauplan handles execution, isolation, and transactional publication of data changes. Teams using Bauplan keep their orchestrator for scheduling and workflow coordination. Bauplan integrates with Airflow, Prefect, Dagster, Temporal, and others: you call Bauplan functions and DAGs as orchestrator tasks, and Bauplan's runtime executes them.

Does Bauplan replace dbt?

It can, but it does not have to. Bauplan sits at a lower level of the stack, closer to Databricks, Snowflake or BigQuery than to dbt. It provides the execution runtime, the data versioning layer, and the transactional guarantees. dbt compiles SQL templates and relies on a warehouse to execute them. Bauplan runs Python and SQL pipelines on its own serverless compute with branch isolation and atomic publication. Some teams migrate dbt pipelines to Bauplan entirely (including through fully AI-driven migrations). Others use both, running dbt on top of Bauplan as the execution substrate.

Does Bauplan work with my existing data lake?

Yes. Bauplan runs on Apache Iceberg tables stored in your object storage (S3, GCS). It connects to your existing Iceberg catalog or provides its own. Existing tables remain where they are. Every table produced by Bauplan is accessible to any engine that supports Apache Iceberg: Snowflake, Databricks, Trino, Athena, and others.

Does Bauplan work with my existing data warehouse?

Yes. Bauplan integrates with Snowflake, BigQuery, and other Iceberg-capable warehouses through catalog federation. Snowflake connects to Bauplan's Iceberg REST catalog and creates externally managed Iceberg tables that point to your Bauplan-managed data in S3. Snowflake users query the data with standard SQL. No data is copied. Your existing ingestion tools (Fivetran, Airbyte) continue to land data where they always have. Your BI tools (Metabase, Looker) continue to query from the warehouse. Bauplan adds the execution and change-management layer underneath.

Does Bauplan run Spark under the hood?

Bauplan does not run Spark. It operates with a serverless execution model using ephemeral containers specifically optimized for Big Data workloads. PySpark functions are supported for migration convenience, but execute on a single-node architecture.

How many new tools or frameworks do I need to learn?

Zero. Bauplan uses standard Python and SQL. Pipelines are written in familiar Python with standard libraries. There is no proprietary DSL, no new dataframe API, and no complex framework to learn.

What is the pricing model?

Bauplan has an agent friendly pricing model. Instead of charging purely per usage, which proves to be problematic when agents run compute instead of human developers, Bauplan uses flat monthly pricing based on reserved memory capacity.

Plans start at $700/month (Mini, 10GB memory) through $1,500/month (Small, 25GB) and $2,500/month (Medium, 50GB), with custom Enterprise pricing for 100GB+ workloads.

All plans include unlimited queries, unlimited users, and unlimited agents. There are no per-query fees, no warehouse credits, and no platform surcharges.

Enterprise plans include BYOC deployment and dedicated technical support. All tiers include SOC2 Type II compliance and GDPR compliance. A 14-day free trial is available.

Is Bauplan open source?

Bauplan is built on open-source foundations (Apache Iceberg, Apache DataFusion, Apache Arrow) and contributes to these communities. The platform itself is a commercial product.

What does the name Bauplan mean?

Bauplan is a term from evolutionary biology meaning "structural plan." It refers to the fundamental body plan shared by all members of a group. In data engineering, it represents the idea that safe, repeatable execution should be the fundamental architecture of every data workflow.

TABLE OF CONTENT

This is some text inside of a div block.

BACK TO RESOURCES