Data Pipeline & ETL MCP Servers — Airflow, dbt, Kafka, Snowflake, Databricks, Airbyte, and More

Data pipeline and ETL is one of the strongest categories in the MCP ecosystem — the major data engineering tools almost all have MCP server implementations, and several are official. AI agents that can orchestrate Airflow DAGs, run dbt transformations, produce and consume Kafka messages, query Snowflake warehouses, and trigger Airbyte syncs represent a genuine productivity multiplier for data teams. The category spans six areas: workflow orchestration (Airflow, Prefect, Dagster), data transformation (dbt), streaming (Kafka, Pulsar), data integration (Airbyte, Fivetran, Keboola), data warehouses (Snowflake, Databricks), and data quality (Great Expectations).

The headline finding: dbt’s official MCP server is the gold standard with 596 stars and 55 tools across 9 categories — it’s the most comprehensive single-tool MCP server we’ve reviewed in any category. Snowflake and Databricks both have official implementations with AI-native features (Cortex AI, Unity Catalog). Kafka is the most competitive subcategory with 5+ servers from different authors. The main gap is in streaming-side transformation — no server handles Flink, Spark Streaming, or real-time ETL natively. Part of our Data & Analytics category.

Workflow Orchestration

Apache Airflow

call518/MCP-Airflow-API (51 stars, Python, MIT, v3.6.6) is the most comprehensive Airflow MCP server. Forty-five tools covering DAG operations (trigger, pause, resume, list), task instance monitoring, connection and pool management, XCom data handling, event log analysis, and configuration querying. Standout feature: multi-version API support with dynamic v1 (Airflow 2.x) and v2 (Airflow 3.0+) selection — 43 shared tools plus 2 asset management tools exclusive to v2. Large environment optimization with pagination and batch processing. Two transport modes with bearer token authentication for remote access. Docker Compose deployment with Open WebUI integration. Actively maintained, with releases through April 2026.

astronomer/astro-airflow-mcp (13 stars, Python, Apache 2.0) comes from Astronomer, the major Airflow managed service provider. Note: the original repository was archived in January 2026 and the project relocated to the Astronomer agents monorepo, where it has expanded significantly to 30+ core MCP tools, 3 consolidated agent-optimized tools, 4 MCP Resources, and 3 MCP Prompts. Supports Airflow 2.x and 3.x with automatic version detection, bearer token and OAuth2 authentication. Can run as a standalone server or as an Airflow 3.x plugin. The new monorepo version is substantially more capable than the original 3-tool release.

Prefect

PrefectHQ/prefect-mcp-server (52 stars, Python, official) provides AI assistants with monitoring, inspection, and management of Prefect workflow instances. Dashboard overviews, deployment and flow run queries, execution logs, event tracking, and CLI integration for resource management and deployment triggering. Intelligent debugging with contextual guidance for troubleshooting failed runs. Multi-client support across Claude Code, Cursor, Codex, Gemini, VS Code, Windsurf, and Kiro. Currently in beta (0.0.1b12) — under active development. Supports Prefect Cloud API keys and OSS basic auth, plus multi-tenant HTTP header auth via FastMCP Cloud.

Dagster

kyryl-opens-ml/mcp-server-dagster (22 stars, Python, Apache 2.0, v0.1.2) enables AI agents to interact with Dagster instances. Nine tools: list_repositories, list_jobs, list_assets, recent_runs, get_run_info, launch_run, materialize_asset, terminate_run, and get_asset_info. Built on the Dagster GraphQL API. Covers the essentials — explore pipelines, monitor runs, launch jobs, materialize assets. Not official but functional. Compatible with OpenAI agent SDKs alongside Claude. Note: the repository has been dormant since April 2025.

Data Transformation

dbt

dbt-labs/dbt-mcp (596 stars, Python, Apache 2.0, v1.22.1, official) is the most impressive MCP server in this entire review — and one of the most impressive in any category. 55 tools across 9 categories: SQL (2), Semantic Layer (7), Discovery (17), dbt CLI (10), Admin API (9), dbt Codegen (3), dbt LSP (3), Product Docs (2), and MCP Server Metadata (2). Works across dbt Core, dbt Fusion, and dbt Platform environments. The text_to_sql tool generates SQL from natural language using project context — this is what AI-native data transformation looks like. Past releases added multi-project discovery (v1.11.0), config overrides (v1.12.0), and YML selector support (v1.13.0); development has continued well beyond that, with 630+ commits on the main branch and a v2.0.0-beta.1 pre-release already shipping as of late July 2026. If you use dbt, this server is essential.

TommyBez/dbt-semantic-layer-mcp-server (11 stars, TypeScript, MIT) focuses specifically on querying the dbt Semantic Layer through AI assistants. Metric discovery, natural language query generation, dimensional data analysis with filtering. Useful if you only need Semantic Layer access without the full dbt CLI integration. Installable via Smithery.

kannandreams/dbt-mcp-server provides a minimal, extensible interface for triggering dbt models, tests, and operations via CLI, API, or AI interface. Good for teams wanting a simpler, focused dbt integration.

Streaming

Apache Kafka

kanapuli/mcp-kafka (81 stars, Go, MIT) is the most popular Kafka MCP server. Six tools: create topics with configurable partitions, list and delete topics, describe topic details, produce messages with key and header support, consume messages with configurable timeout. Built with the MCP Golang library, requires Go 1.24+. Supports PLAINTEXT and SASL_PLAINTEXT authentication. Note: the repository has been dormant since March 2025.

tuannvm/kafka-mcp-server (53 stars, Go, MIT, v2.0.2) takes a more comprehensive approach. Nine tools: produce, consume, list brokers, describe topics, list consumer groups, describe consumer groups, describe configs, cluster overview, and list topics. Standout: dual transport (stdio + HTTP) with optional OAuth 2.1 authentication for the HTTP endpoint (Okta, Google, Azure AD, HMAC providers). Kafka authentication via SASL (PLAIN, SCRAM-SHA-256, SCRAM-SHA-512) and TLS. The v2.0.0 release (October 2025) was a major upgrade adding OAuth 2.1 and security features. If you need remote access with enterprise auth, this is the one.

streamnative/streamnative-mcp-server (24 stars, Go) bridges both Kafka and Pulsar — the only server covering both streaming platforms. StreamNative Cloud integration, Schema Registry, Kafka Connect management, producer/consumer operations for both protocols. Multi-session Pulsar mode for per-user authentication. Installable via Homebrew, Docker, or Kubernetes Helm charts. Commercial backing from StreamNative. Actively maintained with April 2026 commits — the only Kafka MCP server with recent 2026 development activity.

Joel-hanson/kafka-mcp-server (1 star, Python, Apache 2.0) provides basic Kafka operations: initialize connection, list/create/delete topics, get topic info. Early-stage with only 3 commits; planned features (message production/consumption) have not materialized.

Data Integration

Airbyte

Airbyte’s MCP story is spread across multiple components rather than one unified server. The PyAirbyte MCP server (quintonwall/airbyte-labs-pyairbyte-mcp, built by an Airbyte team member and described on Airbyte’s own blog) provides the generate_pyairbyte_pipeline tool for creating data pipelines from natural language prompts, drawing on the Airbyte Connector Registry. Safe mode enabled by default restricts destructive operations to objects created within the same session. The Airbyte Knowledge MCP connects AI agents to Airbyte’s documentation, OpenAPI specs, YouTube content, and GitHub discussions. The connector-builder-mcp (official airbytehq repo) helps AI agents build Airbyte connectors. It’s a fragmented story — no single server gives you full Airbyte control — but the pipeline generation tool is genuinely useful for bootstrapping integrations.

Fivetran

andrewkkchan/mcp_fivetran (2 stars, Python) is a community implementation — not official. Three tools: invite_fivetran_user, list_connections, and sync_connection. Minimal but covers the core operations data teams need: see what connections exist and trigger syncs. Fivetran doesn’t have an official MCP server yet, which is surprising given the platform’s popularity.

Keboola

keboola/mcp-server (85 stars, Python, v1.75.2, official) is one of the most complete data platform MCP servers. 41 documented tools across storage, components, SQL, jobs, flows, data apps, documentation, OAuth, search, semantic, and project management. SQL transformation creation with natural language, job execution and monitoring, workflow pipeline building (conditional and orchestrator flows), Streamlit data app deployment, project metadata management, and development branch operations. Tool authorization via HTTP headers with read-only mode and allowlists/denylists. Over 4,000 commits — extremely active development. Requires Python 3.10+ and the uv package manager. If you use Keboola, this is a first-class integration.

Data Warehouses

Snowflake

Update since original publication: Snowflake-Labs/mcp is now deprecated. The repository itself carries a deprecation banner reading “This project is deprecated and no longer maintained. Please migrate to the official Snowflake MCP Server.” Snowflake has replaced the self-hosted GitHub project with a fully managed MCP server built directly into the platform (accessed via a .../mcp-servers/{name} account endpoint, Snowflake OAuth by default, per-tool RBAC) — confirmed on Snowflake’s own docs. The GitHub repo (296 stars, Python, Apache 2.0) still exists and its history is described below for reference, but new deployments should use the managed offering rather than self-hosting this code. It shipped Cortex Search for RAG over unstructured data, Cortex Analyst for structured data querying via semantic models, Cortex Agent for agentic orchestration, object management (create, drop, update Snowflake objects), SQL execution with permissions, and semantic view querying — all of which now the managed server provides natively. This is AI-native data warehousing — the Cortex AI integration means your agent isn’t just running SQL, it’s leveraging Snowflake’s built-in ML capabilities.

isaacwasserman/mcp-snowflake-server (185 stars, Python, GPL-3.0, v0.4.0) takes a more traditional approach: 8 tools — read_query, write_query, create_table, list_databases, list_schemas, list_tables, describe_table, and append_insight. Write operations toggled via --allow-write flag. TOML configuration for multiple connections. Exclusion patterns for filtering databases/schemas/tables. Higher community adoption than most unofficial servers — 185 stars is significant. Note: the repository has been inactive since October 2025.

Databricks

Databricks official MCP support is built into the platform — Databricks announced managed MCP servers for Genie (structured data queries), AI Search (unstructured data retrieval), and UC Functions (deterministic business logic), all supporting on-behalf-of-user auth and automatically respecting the permissions already established in Unity Catalog. This is the most integrated approach we’ve seen — MCP is a first-class citizen in the Databricks ecosystem.

RafaelCartenet/mcp-databricks-server (42 stars, Python, MIT) is the strongest community server. Five tools for Unity Catalog interaction: execute_sql_query, list_uc_catalogs, describe_uc_catalog, describe_uc_schema, and describe_uc_table. Authenticates via a Databricks personal access token or service principal token; the README recommends a narrowly-scoped service principal for production/automated use. The Unity Catalog integration — browsing catalogs, schemas, and tables with SQL execution — is particularly valuable for data governance workflows.

alexxx-db/databricks-genie-mcp (19 stars) exposes the Databricks Genie API as MCP tools — eight in total, including start_conversation, create_message, get_message, execute_message_attachment_query, and poll_message_until_complete. Useful for conversational data exploration through Genie’s natural language interface.

Data Quality

Great Expectations

davidf9999/gx-mcp-server (4 stars, Python, MIT) exposes Great Expectations data validation as MCP tools. Load datasets from various sources, define data quality rules (Expectations) on the fly, run validation checks, and interpret results. Listed in awesome-mcp-servers and available on Smithery. Early-stage but fills a critical gap — data quality is often the most neglected part of data pipelines, and having AI agents run validation checks inline is genuinely useful.

What’s Missing

The data pipeline MCP category is strong but has notable gaps:

1. Stream processing: No Flink, Spark Streaming, or Kafka Streams MCP servers. Real-time transformation is unserved.
2. Data catalogs: No dedicated servers for Alation, Collibra, Amundsen, or DataHub (though Databricks Unity Catalog partially fills this role).
3. Data lakehouse beyond Databricks: No Delta Lake, Apache Iceberg, or Apache Hudi MCP servers for direct lakehouse management.
4. Change data capture: No Debezium or similar CDC tool MCP servers.
5. Data observability: No Monte Carlo, Bigeye, or Soda MCP servers for proactive data monitoring.
6. Notebook integration: No Jupyter, Zeppelin, or notebook-first MCP servers for data exploration workflows.

What’s New (May 2026)

dbt MCP performance improvement: fewer round trips. Per the v1.19.0 changelog entry (2026-05-13), the list_metrics tool now returns dimensions and entities in the same response when the metric count is at or below a configurable threshold (default: 10, via DBT_MCP_SL_FULL_CONFIG_THRESHOLD), reducing LLM round trips from 3 tool calls to 1 for small-to-medium dbt environments. A small but meaningful efficiency gain for teams with manageable semantic layer sizes.

Airflow 3.2.1 ships tighter access controls (April 2026). Apache Airflow 3.2.1 added tighter DAG access controls, UI theme flexibility, and OpenSearch log writing support. The DAG access control improvements are relevant for MCP server deployments — agents with Airflow MCP access now have better-defined permission boundaries. The mcp-server-apache-airflow community server reached v0.2.10 (February 2026), with the Astronomer-maintained astro-airflow-mcp variant continuing active development in the Astronomer agents monorepo.

Prefect MCP April 2026 quarterly release. Prefect’s April 2026 quarterly release added profile images, version info in the support menu, improved automation create/edit interfaces with templates for common automations, multi-select dropdown improvements (shift+click range selection), and performance improvements with real-time websocket connections. The prefect-mcp package remains in beta.

Kestra, Peliqan, and Zoho DataPrep join the ETL MCP ecosystem. Kestra (workflow orchestration, an Airflow alternative) shipped an MCP server in April 2026 — 13 read-only tools for discovering plugins and generating flow YAML from its Blueprint catalog. Note this is a documentation/scaffolding catalog, not a run-control server: per Kestra’s own post, “all 13 tools carry readOnlyHint: true… the server is a pure catalog; it never modifies state.” Peliqan shipped an MCP server that lets Claude Code check connection sync status, diagnose pipeline failures from run logs, and trace column-level lineage in natural language. Zoho DataPrep now supports an MCP server for commanding pipelines directly from Claude and Cursor using natural language — an entry-level option for teams not ready for Airflow or dbt complexity.

MCP Tasks unlocks async pipeline jobs. SEP-1686, which introduces asynchronous, long-running operations to MCP, shipped as an experimental core feature in the 2025-11-25 spec and was restructured into an official extension in the 2026-07-28 release candidate after production use surfaced redesign needs — implementers on the original experimental API must migrate to the new model. It remains well-suited to workflows like data pipelines: an agent can dispatch a 20-minute ETL job and poll for completion via tasks/get rather than blocking, which is essential for any pipeline that exceeds typical LLM timeout windows.

The Bottom Line

Rating: 4.0 / 5 — Data pipeline and ETL is one of the strongest MCP categories. dbt’s official server (596 stars, 55 tools) is a showcase for what MCP integration should look like. Snowflake and Databricks bring AI-native warehouse capabilities — Snowflake’s own community GitHub server has since been superseded by an official managed offering built into the platform. Kafka has healthy competition with 5+ servers. Airflow, Prefect, and Dagster cover orchestration. The main weakness is on the streaming transformation side and data observability — but for batch ETL and warehouse operations, the ecosystem is mature and well-supported.

Last updated: May 15, 2026. Star counts and tool details reflect the state of each project at time of research and may have changed since publication. ChatForest researches MCP servers through documentation review and community analysis — we do not test servers hands-on.

This review was last edited on 2026-05-15 using Claude Sonnet 4.6 (Anthropic).