Skip to content
Upstream

GUIDE, read 10 Nov 2026

Lineage is usable by an agent when it is served through an authenticated tool that traverses the graph at column level

The matrixAll guides

An agent needs a query interface, not a diagram

A lineage graph drawn for people is a picture. An agent cannot read a picture. It needs a call it can make, such as get upstream columns of this field, and a response it can reason over. Most tools now offer this through an MCP server, an SDK or an API.

Three properties decide whether the answer is useful. First, depth: can the call return several hops, or only direct neighbors? Second, grain: does it return columns or only tables? Third, scope: does the answer respect what the asking user is allowed to see? A fourth, which people forget, is whether the call can change anything. A lineage question should never need write access, and a server that bundles read and write tools needs a way to switch the writes off.

Depth and grain: which tools return paths and columns

DataHub's MCP server traces lineage at table and column level, upstream or downstream, across multiple hops. Its get_lineage_paths_between tool returns exact lineage paths including intermediate transformations and SQL query information. Collibra's server exposes get_lineage_upstream, get_lineage_downstream, get_lineage_entity and get_lineage_transformation, plus search tools for lineage entities and transformations. Collibra also says its server can trace a measure back to its underlying physical columns and tables.

SYNQ's remote MCP server lists get_lineage, get_upstream_dependencies, get_downstream_dependencies, get_entity_impact and get_code. Alation's agent SDK has a lineage tool and a LangChain adapter. Secoda's server exposes entity_lineage and glossary tools, along with run_sql.

Foundational hosts a remote MCP server that gives AI tools end-to-end lineage, usage, code definitions and ownership. Select Star says its MCP server gives agents access to metadata, lineage and semantic models through one API. Coalesce Catalog's server has 16 tools that search metadata, list table columns and traverse lineage.

What an agent does with lineage

Agents use lineage in a few repeatable ways. A coding agent about to edit a model asks what is downstream. An analyst agent asked why a number moved asks what feeds it, and what changed upstream. A governance agent asked to find every report touching a sensitive column walks the graph outward from that column. Each of these is a multi-hop traversal. An interface that returns only direct neighbors forces the agent to call it repeatedly, and each extra call is a chance to stop early or to be rate limited.

The size of the answer matters as much as its depth. A column feeding forty dashboards returns forty rows, and the agent must fit them in its context. Interfaces that return paths between two named assets, as DataHub's get_lineage_paths_between does, give a smaller and more direct answer than a full downstream dump.

Scope: whose permissions the answer follows

An agent that can see more than the user who asked is a leak. The evidence shows several designs.

  • Per-user sign-in. Atlan says its MCP supports OAuth per-user authentication so each call runs as the signed-in user, and that results are scoped to the identity's persona and role permissions. Coalesce Catalog uses OAuth sign-in per user. Euno's MCP uses the signed-in user's active persona, and Euno says it checks authentication and persona permissions on the server for every call. SYNQ's consent screen offers only what the user's role allows.
  • Per-user tokens. DataHub says that in a shared deployment each MCP user authenticates with their own token, so permissions and search results stay per user.
  • Role-filtered tools. Alation's remote MCP server filters out tools the current user cannot use. Collibra says some tools require extra OAuth scopes and report a permission error when the user lacks them.
  • Not found. For Select Star's MCP and for IBM's Manta pages, per-user scoping is not documented on the pages read (none found on 10 Nov 2026). IBM states only that agents can act within approved boundaries.

Read-only controls

A lineage agent mostly needs to read. Tools that let it write add risk. Atlan offers an optional read-only mode that removes write, admin and lifecycle tools but keeps search and lineage. DataHub marks tools with read-only and destructive hints and keeps mutation tools off unless TOOLS_IS_MUTATION_ENABLED is set to true. SYNQ's read-only tier makes no changes, with write, governance and deploy as separate opt-in tiers. Coalesce Catalog's MCP is off by default and an admin turns it on, and its write tools need the Contributor role or above.

What each tool offers, side by side

The table lists the agent interface each tool documents for lineage, as read on 10 Nov 2026.

ToolInterface documentedScoping documented
DataHubMCP server, multi-hop table and column lineagePer-user token in shared deployment
AtlanHosted MCP, optional read-only modeOAuth per user, persona and role scoped, read-only mode
CollibraMCP with lineage upstream, downstream, entity and transformation toolsOAuth with scopes
AlationRemote MCP and agent SDK with a lineage toolTools filtered by user
SYNQRemote MCP with lineage, dependency and impact toolsRole tiers on consent screen
SecodaMCP with entity_lineageRespects workspace permissions
Coalesce CatalogMCP (16 tools), off by defaultOAuth per user
FoundationalRemote MCP (Streamable HTTP)Tied to a user account through browser sign-in
Select StarMCP with metadata, lineage and semantic modelsNone found on 10 Nov 2026
EunoHosted MCP endpoint with OAuth sign-inPersona of the signed-in user
OpenLineage and MarquezMarquez lineage APIMCP or agent API: none found on 10 Nov 2026

Open lineage servers are an API, not an agent interface

Marquez exposes a lineage API so metadata can be queried programmatically, for example to enrich catalogs and data quality systems, per the Marquez site. That is useful, and an agent can call it. But no MCP server or agent-specific API from OpenLineage or Marquez was found on 10 Nov 2026, and no permission model was found either. A team using Marquez would write its own tool wrapper and its own access control.

How to test an agent against lineage

Select Star says its MCP server works with Cursor, Claude Desktop and VS Code GitHub Copilot. These clients speak the same open protocol, so client support is rarely the deciding factor. What decides is what the tools return and whose permissions they respect.

Ask the agent three questions on your own data, using a test user whose role is narrower than an admin's. Which dashboard fields depend on this warehouse column? Which transformation changes it on the way? What can a user with the analyst role not see? A server that answers the first from a table-level graph, or answers the third with data the role should not see, fails. The glossary defines the terms used here, and the tool pages list the evidence behind each row.

Sources

  1. DataHub: docs.datahub.com/docs/features/feature-guides/mcp, read 10 Nov 2026
  2. Collibra: productresources.collibra.com/docs/collibra/latest/Content/ModelContextProtocol/, read 10 Nov 2026
  3. SYNQ: docs.synq.io/scout/mcp.md, read 10 Nov 2026
  4. Alation: github.com/Alation/alation-ai-agent-sdk, read 10 Nov 2026
  5. Atlan: docs.atlan.com/product/capabilities/atlan-ai/references/mcp-security, read 10 Nov 2026
  6. Secoda: docs.secoda.co/features/ai-assistant/secoda-mcp-server.md, read 10 Nov 2026
  7. Foundational: docs.foundational.io/en/articles/11194236-use-foundational-mcp-with-your-ai-tool, read 10 Nov 2026
  8. Select Star: www.selectstar.com/product/mcp-for-data, read 10 Nov 2026
  9. Coalesce Catalog: docs.coalesce.io/docs/catalog/developer/mcp-integration.md, read 10 Nov 2026
  10. Euno: docs.euno.ai/ai/mcp-setup.md, read 10 Nov 2026
  11. OpenLineage: marquezproject.ai, read 10 Nov 2026
  12. OpenLineage: openlineage.io/docs, read 10 Nov 2026

Last reviewed 10 Nov 2026.