GUIDE, read 10 Nov 2026
An MCP server for data exposes up to five kinds of tools, and each server's tool list shows which kinds it actually ships
Query tools run SQL or a modeled query and return rows
A query tool sends a statement or a structured request to an engine and returns rows. It is the kind of tool most engineers expect first, and the servers split on what the statement is. Google's BigQuery reference lists two SQL tools: execute_sql_readonly, which Google says is restricted to SELECT statements, and execute_sql, which can run INSERT, UPDATE, DELETE and CREATE statements. The Amazon Redshift MCP Server from AWS Labs lists one, execute_query.
Some catalog servers also carry a query tool. Per Atlan's tools page, query_assets runs read-only SQL against a connected warehouse (Snowflake, Redshift, BigQuery or Databricks) and returns up to 100 rows. Secoda's docs describe run_sql as executing SQL against connected warehouses. Databricks documents a dbsql MCP that runs SQL under the caller's Unity Catalog and warehouse permissions; per Databricks, dbsql sits behind the Unity Gateway beta.
A modeled query is the other form. Per Cube's docs, runQuery runs Cube SQL against views whose names and members come from searchDataModel, so the agent writes against the semantic model and not against raw tables.
Metadata tools describe schemas, assets and owners
A metadata tool answers what exists and what it means without returning business rows. The Redshift server lists list_clusters, list_databases, list_schemas, list_tables and list_columns, which walk from cluster to column. BigQuery's remote server lists list_dataset_ids, get_dataset_info, list_table_ids and get_table_info.
The depth differs. Per the community mcp-snowflake-server README, describe_table returns column names, types, nullability, defaults and comments, so table and column comments reach the agent. Per Tableau's docs, get-datasource-metadata returns table relationships, field descriptions, roles and types, calculation strings and parameters for a data source. The DataHub MCP Server lists list_schema_fields and get_entities, and its README says agents can discover domains, owners, tags and glossary terms.
Per Collibra's README, discover_business_glossary and get_column_semantics reach business terms connected to a column. For these catalog tools the content comes from what the catalog holds, so the README does not list the warehouses behind it (none found on 10 Nov 2026 in the Collibra README).
Pipeline and observability tools add a state dimension to metadata. Sifflet's search_asset filters by asset type TABLE_AND_VIEW, PIPELINE, DASHBOARD or ML_MODEL. Airbyte's describe_connection returns a connection's streams, schedule and status, and discover_source_catalog is a catalog tool at stream level. Per Monte Carlo's docs, the extended toolset adds get_use_cases and get_use_case_tables for critical-table groupings.
Lineage tools trace what feeds an asset and what depends on it
Lineage tools return a graph instead of a row set. Per the dbt Labs README, get_lineage returns a bounded graph with type, depth and direction filters, and column-level lineage comes from fusion.get_column_lineage on dbt Platform and get_column_lineage locally. Per DataHub's README, get_lineage_paths_between returns exact paths between two assets or columns, including intermediate transformations and SQL query information.
Per Collibra's README, get_lineage_transformation returns the details and logic of a specific transformation. Sifflet's server source lists get_downstream_assets_of_asset, which returns downstream assets by URN. Select Star's page lists traverse_lineage, and the Select Star MCP Server page says nothing on column-level depth (none found on 10 Nov 2026). Per Euno's docs, get_upstream_lineage and resource_impact_analysis work on one resource URI.
Warehouse servers tend to publish none. The Redshift server (8 tools), the BigQuery remote server (9 tools), the Snowflake-managed server and the Databricks-provided MCPs list no lineage tool; each page was read on 10 Nov 2026 and the finding is recorded as none found.
Semantic layer tools return governed definitions instead of raw tables
A semantic layer tool hands the agent metrics, dimensions and business definitions, so a question like revenue by region resolves to a defined measure. Per the dbt Labs README, the Semantic Layer tools list metrics, dimensions, entities and saved queries, and query_metrics runs a metric query. Per Cube's docs, searchDataModel searches views and their measures and dimensions by similarity.
Snowflake's docs say the Cortex Analyst tools on the Snowflake-managed server use semantic views, which hold business definitions, and that the server accepts semantic views only, not semantic models. Databricks says Genie One resolves business terms, metric definitions and table relationships through Genie Ontology, which it describes as a governed semantic layer. Per Tableau's docs, the server serves Pulse metric definitions so metrics are answered from governed definitions. Per Omni's docs, queries run against modeled topics.
The vendors describe these layers in their own words, and the layers are not interchangeable: a dbt metric, a Cube view and a Snowflake semantic view are defined in different places and queried through different tools.
Write actions change data, models or metadata
Write tools are the fifth kind, and they sit inside otherwise read-oriented servers. The dbt Labs README warns that its CLI tools (build, run, test) can modify data models and warehouse objects. Collibra's edit_asset sets attributes such as Definition and assigns Steward and Owner roles, per its README. DataHub's optional tools add tags, terms and owners. Hex lists create_cell, update_cell and delete_cell, and Airbyte lists deploy_connector and permanently_delete_connector.
Which servers ship writes, and how each documents confirmation, is covered in the guide on read versus write. This guide only records that the kind exists and that a tool list shows it by name.
Write tools are not limited to warehouse data. Per Euno's docs, its catalog tools create, update and delete glossary terms, metadata tags and workflows. ThoughtSpot's tool definitions include createLiveboard, create_dashboard and modify_dashboard, which create or change content inside the BI tool, and Cube's tools create reports and workbooks.
A tool list shows which kinds a server ships
The table puts a sample of servers side by side by the kinds named in their own tool lists or docs. A dash means none found on 10 Nov 2026 in the recorded facts. The table is a reading aid and not a measure of how many tools a server has; the Context Weight Index covers tool-definition size separately.
Cells marked none found mean the tool list or docs page read on 10 Nov 2026 named no tool of that kind. They do not say the capability is unavailable through other means, such as the engine's own features.
| Server | Query | Metadata | Lineage | Semantic layer | Writes |
|---|---|---|---|---|---|
| Amazon Redshift MCP Server (AWS Labs) | execute_query | list_clusters to list_columns | none found | none found | read-write mode, opt-in |
| BigQuery remote MCP server (Google Cloud) | execute_sql_readonly, execute_sql | get_table_info, get_dataset_info | none found | none found | execute_sql |
| Collibra MCP Server | none found | discover_data_assets, get_column_semantics | get_lineage_transformation | get_table_semantics | edit_asset, create_asset |
| Cube MCP server | runQuery | searchDataModel | none found | views, measures, dimensions | writeDataModelFile, createReport |
| DataHub MCP Server | none found | search, list_schema_fields | get_lineage | none found | add_tags, add_owners (off by default) |
| dbt MCP Server | execute_sql, text_to_sql | get_all_models | get_lineage, get_column_lineage | list_metrics, query_metrics | build, run, trigger_job_run |
Sources
- BigQuery MCP tools reference, read 10 Nov 2026
- AWS Labs Redshift MCP README, read 10 Nov 2026
- Atlan MCP tools page, read 10 Nov 2026
- Secoda MCP server docs, read 10 Nov 2026
- Databricks managed MCP page, read 10 Nov 2026
- Cube MCP server docs, read 10 Nov 2026
- Community mcp-snowflake-server README, read 10 Nov 2026
- Tableau MCP docs intro, read 10 Nov 2026
- DataHub MCP README, read 10 Nov 2026
- Collibra MCP README, read 10 Nov 2026
- dbt MCP README, read 10 Nov 2026
- Sifflet MCP server source, read 10 Nov 2026
- Select Star MCP docs, read 10 Nov 2026
- Euno MCP setup docs, read 10 Nov 2026
- Snowflake-managed MCP docs, read 10 Nov 2026
- Omni MCP tools, read 10 Nov 2026
- Hex MCP server docs, read 10 Nov 2026
- Airbyte MCP tools, read 10 Nov 2026
Last reviewed 10 Nov 2026.