Lineage tooling in practice: DataHub MCP and the lineage-graph plugin

At a glance

Learning objectives

• Wire mcp-server-datahub and call its four tools
• Use /impact, /lineage-up and /lineage-down for change analysis
• Choose the right tool for a catalog question versus an impact question

What you need

• [object Object]
• [object Object]
• DATAHUB_GMS_URL and DATAHUB_GMS_TOKEN for the GMS endpoint

Steps

1. Meet the catalog door — an MCP server that exposes a DataHub catalog so an agent can search entities, read lineage and schema, and inspect metadata. (from mcp-datahub )
2. Learn the tool surface — search, get_entities, get_lineage and get_dataset_queries. (from mcp-datahub )
3. Install it — uvx mcp-server-datahub, with the GMS URL and token wired through environment variables. (from mcp-datahub )
4. Meet the plugin — column-level impact analysis over a lineage store, with a change-impact command that lists the downstream consumers a schema change would break. (from plugin-lineage-graph )
5. Learn the command surface — /impact for blast radius, /lineage-up for what feeds an asset, /lineage-down for what depends on it. (from plugin-lineage-graph )
6. Keep one vocabulary — the metadata model explains how to represent ownership and lineage consistently across tools. (from ref-datahub-docs )
7. Scope before sharing — read-only against the catalog, but catalogue metadata can itself be sensitive; use a scoped token and keep the GMS endpoint internal. (from mcp-datahub )

Check yourself

• get_lineage answers "what feeds this table"; /impact answers "what breaks if this column changes" — different questions.
• Both tools read the same metadata model, so an answer from one is not contradicted by the other.

Worked example

uvx mcp-server-datahub
# required env
#   DATAHUB_GMS_URL=<url of the DataHub GMS endpoint>
#   DATAHUB_GMS_TOKEN=<personal access token>

/impact <asset|column>   # downstream consumers a change would break
/lineage-up <asset>      # what feeds this
/lineage-down <asset>    # what depends on this

Sourced from

• DataHub MCP Server
• Lineage Graph
• DataHub Documentation

Security scan

Last sweep 2026-10-24: 0 finding(s), 0 marked as examples.

Common gotchas

• Catalog metadata is sensitive even when read-only — PII tags and internal table names ride along; use a scoped token.
• get_dataset_queries and get_lineage answer different questions — searching for the wrong one wastes a turn.
• The plugin's /impact walks the same store the MCP server reads — configure both against the same GMS.

Production considerations

• Keep the GMS endpoint internal and hand each integration its own scoped token.
• Treat the four MCP tools as the stable contract — scripts that call them survive catalog UI redesigns.

Diagram

lineage tooling
   |
   |- mcp-datahub ──▶ search | get_entities | get_lineage | get_dataset_queries
   |        (uvx mcp-server-datahub, DATAHUB_GMS_URL + TOKEN)
   |
   |- plugin-lineage-graph ──▶ /impact | /lineage-up | /lineage-down
   |
    --> both read the same metadata model (ownership + lineage)

Registry stats