Your agent isn't wrong.Its context is

It reached your warehouse fine. It just couldn't tell which table was certified, which definition finance uses, or which columns it was never allowed to see, so it guessed and answered anyway. DataHub is the governed context layer that answers all three at query time.

Book a demo

Prefer to look around first?
Take the self-guided product tour

analytics agent · text-to-sql
“What was Q3 net revenue from enterprise accounts?”
$48.2M

Stated with full confidence. No source named, nothing to check it against.

What the agent knew when it answered
Same model, same question. Switch the retrieval path to see what changes.
Trusted by modern data teams

Models don't lack intelligence. They lack governed context

Retrieval hands the agent a table name and a column list. The five things that decide whether an answer is safe to act on aren't in the schema at all, so the agent infers them.

What's missing
Ungoverned retrieval
With a governed context layer
Lineagewhere the number came from
The agent sees a table, not the five hops behind it. It can't tell a source of truth from a stale copy.
Column-level lineage with the transforms in between. The answer names its own origin.
Certificationwhether it's approved
Nothing separates a certified table from one deprecated months ago. Both look usable from the schema.
Lifecycle stage travels with the asset, or deprecated assets are scoped out of the agent's view entirely.
Definitionswhich meaning applies
Three teams each maintain a valid definition of the same metric. The agent picks one and never says which.
One governed glossary term, validated by its owner. The response cites the definition it used.
Owner & freshnesswhether it's current
No owner, no SLA, no quality signal. An agent answers just as confidently from a table that stopped loading on Tuesday.
Owner, freshness, and assertion status on the asset. A failing table is known-failing before it reaches an answer.
Access policywhether it's allowed
Governance in a spreadsheet governs nobody. The agent surfaces a PII column because nothing said not to.
The policy engine that governs people governs every agent, enforced before context is returned, not after.

And it's a maintenance problem, not just a collection one. Definitions move as the business moves. A context layer assembled once and owned by nobody decays back into the fragmentation it was built to fix, and agents start returning answers that used to be right.

See what governed retrieval actually returns.

Book a demo

One governed context layer. Every agent pulls the same answer from it

01

Ingest what's scattered

Definitions in dbt, join logic in Looker, metadata in Snowflake, knowledge in Notion. Event-driven sync from 100+ sources, so a change upstream is live rather than queued for the next batch.

Context ingestion
02

Skip the blank page

Hand-documenting context runs 15+ hours per table. DataHub extracts meaning from query logs, dbt projects and dashboards instead, producing candidate definitions in days.

Context intelligence
03

Put a human in the loop

Auto-generated context is a start, not a finish. Domain experts get a review queue to confirm and resolve conflicting definitions. It becomes a recurring task, not a workshop.

Context hub
04

Serve it everywhere

Snowflake Cortex, Databricks Genie, Claude, Cursor, and custom agents read the same validated context. Build the layer once; every surface inherits it.

Context activation
~/.cursor/mcp.json OAuth + DCR
{
  "mcpServers": {
    "datahub": {
      "url": "https://mcp.datahub.com/mcp"
    }
  }
}

Runs where your agents already run

ClaudeCursorChatGPTSnowflake CortexDatabricks GenieWindsurfLangChainGoogle ADKREST / GraphQL

What an agent actually gets back

Real tools on the MCP server, not a roadmap. Each is annotated with MCP-standard hints, so a client can tell read-only calls from ones that change catalog state.

Connect it to your agent in an afternoon.

Book a demo

What the analysts found, before we say anything about ourselves

If you're building the business case, our own numbers are the weakest evidence in the room. Start with the independent research.

~2x
increase in analytics agent accuracy
Miro, DataHub Cloud customer
60x
faster to answer an analytical question
DataHub customers, production
10x
usage of the next most-used agent
Pinterest Analytics Agent
70%
less manual documentation effort
Pinterest

Pinterest turned 400,000 ungoverned tables into its most-used AI agent

Institutional knowledge lived in Slack threads, and analysts spent hours working out which table was current and safe to use. DataHub became the context layer, governing 100,000 curated assets and indexing analyst query history by intent, so the agent grounds answers in what the organization already knew.

Trusted answers in minutes/70% less manual documentation/Read the full story

The questions that actually decide this

The four we hear on every evaluation call. Answering them here is cheaper for both of us than answering them in week three.

Plenty of teams can build v1, and some should. A metadata store, a few connectors and an MCP endpoint is a quarter of work for a competent platform team, and if your estate is small and stable, that may genuinely be enough.

The build cost isn't v1. It's the four things underneath it: connectors for every system you'll add over three years, an event-driven sync that keeps context current at machine speed rather than nightly, a review workflow domain experts will actually use, and a policy engine that governs agents with the same rules as people, with an audit trail. Then you maintain all of it while the business redefines its metrics.

Most context projects don't fail at collection. They fail at maintenance, set up by a collection project that never decided who owned the result. Ask your team what happens in month nine, when the person who built it has moved on. If that answer is solid, build it. The open source project is Apache 2.0 and it's a good foundation, so that's a real option, not a concession.

Core is a genuine open source metadata platform, not a crippled trial. It has 16,000+ community members, 3M+ downloads a month, and a self-hostable MCP server. Here's the split, plainly:

DataHub Core · Apache 2.0
  • Metadata model, catalog and search
  • Table and column-level lineage
  • Business glossary and documentation
  • Assertions, incidents and data contracts as definitions
  • Self-hosted MCP server you run and upgrade
  • Community support, on community timelines
What Cloud adds
  • Managed MCP endpoint with OAuth and SSO, plus scoped per-domain servers
  • Context Intelligence and the SME review queue
  • Agent Context Kit for Snowflake, Databricks, LangChain, ADK
  • Active quality checks and AI anomaly detection
  • RBAC and ABAC, SOC 2 Type II, 99.5% availability SLA
  • In-VPC remote executor, rollout support, a named team

The parts that make context stay correct, namely automated generation, structured expert validation, drift detection and agent-ready delivery, are the Cloud parts. Core gives agents a path to your metadata. Cloud keeps that metadata worth trusting once nobody is watching it closely.

The semantic layer, no. It reconciles them. In most estates finance has a model in the BI tool, product has one in the experimentation stack, marketing a third in attribution. Each is internally consistent; none is authoritative. DataHub ingests those models rather than re-creating them, and resolves which definition an agent should use.

RAG, no. It decides what's worth retrieving. RAG is a technique that runs on top of this. Without a governed layer beneath it, RAG pulls from whatever documents happen to be indexed, including the deprecated wiki page.

The catalog, functionally yes. The architectural difference is who the output is for: a catalog indexes structured metadata and serves it to humans through a portal. A context platform unifies that with the unstructured knowledge that gives it meaning, and serves both humans and agents from one graph.

DataHub operates on metadata and documentation, not on your rows. Ingestion reads schemas, lineage, query logs and the docs you point it at. For sensitive sources, an in-VPC remote executor runs ingestion inside your network, so credentials and data planes stay where they are.

Deployment is cloud-hosted or on-premises, on SOC 2 Type II certified infrastructure with an SLA-backed 99.5% availability commitment, role- and attribute-based access controls, and SSO. Every context access is logged with provenance, which is the record you need when someone asks what an agent acted on three months ago.

Book your custom demo with DataHub

Share a few details about your goals, and we'll be in touch to schedule a personalized walkthrough tailored to your data environment, use cases, and roadmap.

We'll cover:

How DataHub helps teams discover, understand, maintain, and govern their data
Best practices for driving adoption across your org
How to get started quickly, with cloud-native or on-prem deployment

Let's get you the answers and insights you need.

Tell us a little about your goals and we'll be in touch to schedule your walkthrough.

Prefer to look around first?
Take the self-guided product tour

Tailored to your data environment Built around your use cases and roadmap Cloud-native or on-prem, we'll say which fits
Book a demo