DataHub

Name: DataHub MCP Server
Availability: InStock
Author: DataHub

Data & Research34 tools

Search your data catalogue, trace column-level lineage, draft SQL and edit governance metadata from your AI assistant. 34 tools, OAuth sign-in, one coarse application scope.

Verified connector

Listed by Anthropic as a partner connector in its Connectors Directory.

Connection checked by Agentman on 18 August 2026.

Anthropic states this reflects the level of review a connector received, not a security audit.

Connect DataHub via MCP

https://mcp.datahub.com/mcp

Works in any MCP-compatible client. In Agentman Studio it is one click — no config file to edit.

DataHub Tools & Capabilities (34)

Tool names from Anthropic's directory listing. This server requires sign-in, so we could not read tool descriptions or parameter schemas.

Limits

One application scope, datahub:account, with no read/write split. The consent screen offers nothing to decline, and unlike ChartMogul's equally coarse scope this one gates writes. Access control lives entirely in DataHub's role and policy model.
accept_or_reject_proposals lets an agent complete a human review step. DataHub's change proposals workflow exists so owners and stewards approve suggestions; an agent with Editor or Admin privileges can approve them instead. Connect as a Reader if you want the propose-only path.
Accepting a proposal is not undoable through the connector. DataHub records a full log of accepted and rejected proposals per user — an audit trail, not a rollback. Reversing a decision means making a fresh opposing edit.
DataHub Cloud only. The managed endpoint requires Cloud v0.3.12+, OAuth requires v1.0.2+ and mutation tools require v0.3.17+. DataHub Core users self-host and get a materially smaller surface: twelve of the 34 tools, including every glossary, lifecycle and proposal tool, have no implementation in the open-source repository.
The vendor's annotation claim overstates the open-source implementation. DataHub's guide says all tools carry readOnlyHint, destructiveHint and idempotentHint; the repository sets only readOnlyHint, so the writing half of the surface is unlabelled in clients that read those fields. We could not read the hosted server's annotations.
Metadata edits are organisation-wide. Tags, glossary terms, ownership and domains are shared across the workspace. A bulk edit through this connector is visible to everyone using the catalogue, not scoped to the session that made it.
The mutation kill switch is self-hosted only. TOOLS_IS_MUTATION_ENABLED=false disables writes in the open-source server. DataHub documents no equivalent tenant-level toggle for the Cloud endpoint; the Cloud controls are roles, policies and scoped servers.
Scoped MCP servers are private beta. The per-endpoint tool subset — the strongest available narrowing mechanism — is limited to Context Platform private beta customers on DataHub Cloud.
We could not read tool schemas or safety annotations. The endpoint requires authentication, so the 34 count and the names are the directory's, and the behaviour is the vendor's documentation and source. We did read the auth posture and both OAuth descriptors ourselves.
DataHub publishes no MCP-specific rate limit. Its GMS rate limiting guide documents three separate load-protection mechanisms that can each return HTTP 429, but none is described as the MCP server's quota. Treat those figures as context, not as the connector's limit.
"MCP" is overloaded in DataHub's own docs. DataHub uses MCP for Metadata Change Proposal, its internal event-driven write primitive, in pages that have nothing to do with Model Context Protocol. A search of its documentation for "MCP" returns both.

Frequently asked questions

My colleagues have created some proposals, but I'm not seeing this in my Task Center. Why not?

Your privileges are most likely not configured correctly. DataHub's change proposals guide gives exactly this answer, and directs admins to Settings then Permissions to edit the roles and policies so the account holds the review privileges the same page lists. Proposal visibility follows privileges, not the connector.

Is DataHub's MCP server the open-source project or the paid Cloud product?

The endpoint at mcp.datahub.com is DataHub Cloud only. DataHub's guide marks the managed MCP server as available on DataHub Cloud v0.3.12 and later, with OAuth added in v1.0.2 and later. DataHub Core users self-host the open-source server from the acryldata/mcp-server-datahub repository instead.

Can the DataHub MCP server approve a governance proposal on its own?

Yes, if the signed-in account holds the review privileges. The connector lists accept_or_reject_proposals, and DataHub's change proposals guide reserves review to the Manage Tag Proposals, Manage Glossary Term Proposals and related privileges, granted by default to the Editor and Admin roles. Connect as a Reader to remove that ability.

Are metadata edits made through the DataHub MCP server reversible?

The prior values are retained, but no tool restores them. DataHub's aspect versioning design keeps every earlier version of an aspect as an audit trail under version 0 semantics. The 34-name connector listing contains no restore, revert or undo verb, so recovery happens outside the connector.

Which DataHub role does an agent need to edit catalogue metadata?

Editor or Admin. DataHub's roles guide defines Reader as able to read all metadata and unable to edit anything by default, Editor as able to read and edit all metadata without administrative actions, and Admin as able to do everything. Tokens are scoped to the signed-in user.

How do you stop an agent from writing to your DataHub catalogue at all?

Connect with a Reader-role account or a scoped service account. DataHub documents a Reader role that cannot edit by default, and service accounts that carry their own role plus an optional Default View restricting which assets the MCP server can search. Both narrow the connection without touching OAuth.