Google Knowledge Catalog is Google Cloud’s Gemini-powered catalog for data and AI assets, renamed from Dataplex Universal Catalog on April 10, 2026. It holds entries, aspects, a business glossary and data products across a BigQuery, Looker and Vertex AI estate, and serves them to agents over four documented MCP servers. Atlan is the context layer above it, holding what the business agreed rather than what the platform observed.
The search that lands here usually starts somewhere less tidy. A console tab labeled Dataplex. A runbook that says Data Catalog. A Terraform file pointing at dataplex.googleapis.com. Four product names, one service, and the current name has not reached the API, the CLI, the IAM roles or the documentation path. Agents read this metadata directly now, which turns the context layer into an architecture question rather than a naming one.
| Knowledge Catalog | Quick facts |
|---|---|
| Names | Knowledge Catalog from April 10, 2026; before that Data Catalog, Dataplex, Dataplex Universal Catalog |
| API and CLI | dataplex.googleapis.com, gcloud dataplex, docs under /dataplex/docs/ |
| Agent interface | Four MCP servers, two remote and two local, one endpoint read-write |
| Where Google’s docs stop it | “Certain automated semantic extractions might be limited to built-in Google Cloud services” |
| Metering | DCU processing, metadata storage per gibibyte-hour, data tokens for data insights from October 27, 2026 |
What Knowledge Catalog is, and the four names it answers to
Knowledge Catalog is Google Cloud’s catalog for data and AI assets, and Google publishes two definitions of it that differ on reach.
The Knowledge Catalog overview (Google Cloud, last updated 2026-09-15) calls it “a Gemini-powered data catalog that provides universal business context and governance for your entire data estate.” The FAQ (2026-09-15) calls it “an intelligent governance solution for data and AI assets in Google Cloud.” One says entire data estate. The other says in Google Cloud.
Entries, entry groups, aspects and entry links are the vocabulary every quota, migration step and MCP tool runs on, the same vocabulary behind types of metadata for AI agents. Google draws its own scope line, too: the overview and the best practices page (2026-09-15) both say “while Knowledge Catalog supports major third-party systems, certain automated semantic extractions might be limited to built-in Google Cloud services.” Federation registers an asset; extraction, where meaning comes from, is known only because Google Cloud observed it running, narrower than what a data catalog has to do for AI.
Four names, only the last dated. Data Catalog set the tag and tag-template model estates still hold; Dataplex folded cataloging into a lake, zone and asset model; a later consolidation merged Dataplex Catalog with BigQuery metastore. Google gives no announcement date for any of the three, so none is asserted here.
The fourth is dated because Google dated it. Per the release notes (read 2026-09-16): “As of April 10, 2026, Dataplex Universal Catalog is now called Knowledge Catalog. The API, client library, CLI, and Identity and Access Management (IAM) names remain unchanged.” The launch post followed on April 23, and the two dates get conflated constantly. The endpoint is still dataplex.googleapis.com, the roles still read roles/dataplex.catalogAdmin, and the product page is titled “Knowledge Catalog (formerly Dataplex).” A name your own tooling cannot see is a context portability problem before it is a branding one.
What shipped since June, and the migration some estates still owe
Google’s release notes record ten dated changes to Knowledge Catalog between June 4 and September 14, 2026.
Per the release notes (read 2026-09-16 against a page stamped 2026-09-15): agent-ready context bundles, new source connectors, governance workflows, and metric and data-domain imports, most Preview; only Cloud Storage automatic discovery and org-, folder- and project-level lineage control reached GA. Every stage word there is Google’s own.
Two matter more than the rest. Data products are GA per About data products (2026-09-15), close to the shape data contracts for AI take when enforced instead of documented. BigQuery measures are Preview on the metadata overview (2026-09-15), so a team already modeling metrics in Looker runs two definitions in two places, the tension in LookML against a universal semantic layer. Iceberg REST ingestion reaches Databricks Unity, AWS Glue and Snowflake Horizon too, all Preview, and registering a table is not understanding it, which is why Snowflake Horizon Context, Genie Ontology and DataZone against the Glue Data Catalog each keep their own semantics.
Google frames this as grounding on its page for AI agents (2026-09-15). Ten changes in fourteen weeks, almost all adding supply, none changing what the catalog can be authoritative for, and the legacy service it replaced is still mid-shutdown.
The migration the release notes started, and nobody finished
Legacy Data Catalog did not stop cleanly. The deprecations page (2026-09-15) dates deprecation to 2025-02-03 and shutdown to 2026-06-01, Business Glossary on the same date. Two surfaces describe June 1 differently: the Data Catalog banner says the service “will be discontinued on June 1, 2026,” flatly, while the transition guide, still live on 2026-09-15, describes a phased shutdown and still calls Data Catalog “the authoritative source for your custom metadata” during the preparatory phase.
Most readers qualify for the shortcut: no custom metadata, tags, tag templates, or custom entries and entry groups, and setting the default catalog UI to Knowledge Catalog finishes it. Everyone else should hold four facts: regional tag templates become global aspect types, so name collisions become global; private tag templates have no counterpart, because Knowledge Catalog does not support private aspects; Terraform cannot manage custom entries and aspects; and policy tags are explicitly not deprecated, leaving column-level access in its own system beside the aspect model. The quiet one is IAM: propagated aspect types and custom entry groups “don’t automatically inherit original Data Catalog permissions,” so rebuilding them by hand is where estates lose weeks, making role-based access control in a context layer a design surface rather than cleanup. Glossaries move separately, through a dependency-ordered Python import per the migration guide (2026-09-15); the reconciliation, not the script, decides time to value for an AI context platform.
What an agent can actually get: the published limits and four MCP servers
Google publishes hard limits and a four-server agent interface, and together they decide what an agent can retrieve before answer quality is even in question.
The quotas and limits page (2026-09-15) caps a business glossary at 5,000 terms, a project at 200 data products per region with 50 assets each, and search at 500 results in keyword-only mode or 100 otherwise. Two structural limits bite harder than any single number: the catalog “indexes each shard of a date-sharded table as a separate table,” silently multiplying entry count, and natural language search excludes public resources and aspects attached to entry links, so test whether an agent can find cross-asset relationships in English before relying on them. For a human, a result cap is a convenience; for an agent, it is a ranking problem that belongs to the retrieval architecture rather than the catalog’s size, the line between semantic search and keyword search and the reason a context layer and a vector database are not interchangeable.
The agent interface is four servers and five endpoints, because the catalog’s remote server exposes two. Per the MCP servers overview (2026-09-15), local servers run on your own machine over stdio and remote servers run on Google’s infrastructure behind an HTTP endpoint, changing your auth path and audit trail, covered in an MCP architecture deep dive and why MCP matters for AI agents. The separate lineage remote server (2026-09-15) exists because agents ask about provenance and downstream impact when an answer looks wrong, the pattern in MCP for data lineage.
One endpoint can write. Per the MCP reference (stamped 2026-07-13, the oldest document in the set), the core /mcp endpoint exposes search_entries, lookup_context and lookup_entry, all reads, while the /mcp/data-products toolset adds eleven more including create_data_product and update_data_product. An agent that can create a data product can create a definition, so settle precedence before connecting the second server: which layer is authoritative for BigQuery context works that out per object class.
Where the catalog MCP server’s release stage actually stands
Two separate facts make up the answer, and neither is a label. Google announced the catalog remote server as Preview in its release notes under May 25, 2026: “Remote Model Context Protocol (MCP) server support (Preview). Data applications and AI agents can programmatically interact with data products.” As of 2026-09-15 that Preview banner is no longer on the remote MCP server guide, and no GA announcement has followed. The banner came off. The label never went on. The Preview banner still live belongs to the data lineage remote server, not this one, where Google writes “Pre-GA features are available ‘as is’ and might have limited support.”
Google is not vague when it means otherwise: that same May 25 block also says “Data products in Knowledge Catalog is Generally Available (GA),” and a check of the whole release-notes page on 2026-09-16 returns no such wording for any MCP server. Attaching one page’s banner to the other server is how a careful reader gets this wrong.
What Google meters, and the one number no meter reports
Knowledge Catalog bills processing in Data Compute Units and metadata by the gibibyte-hour. A second meter starts charging for Gemini-powered data insights on October 27, 2026, and nothing on either invoice reports whether an agent got its answer.
| What Google meters | Rate, Iowa (us-central1), USD, read in a browser 2026-09-16 |
|---|---|
| Dataplex standard processing, discovery | $0.06 per DCU per unit time |
| Dataplex premium processing, lineage and quality and profiling | $0.089 per DCU per unit time |
| Metadata storage | $0.002739726 per gibibyte-hour, about $2 per GiB per month in Google’s own examples |
| Free tier | 100 DCU-hours per month, Standard SKU only, explicitly not on Premium |
| Data insights, from October 27, 2026 | $3 per million input data tokens, $20 per million output |
Read the free-of-charge list twice: no charge for data organization, security policy application and propagation, auto-ingested Google Cloud technical metadata storage, or API calls, catalog search included. Serving an answer is not a billable event, because serving is not what is metered. The pricing page is titled “Knowledge Catalog (formerly Dataplex) pricing” and its SKU names still read Dataplex, so the rename reached the marketing page and not the invoice.
Now the disagreement, which is the finding. That pricing page describes Gemini-powered features as billing “as part of Gemini in BigQuery or Gemini Code Assist” and names no token meter anywhere. The product’s own FAQ, last updated 2026-09-15, says under a heading of Data insights billing that “Active billing for data insights starts on October 27, 2026” and that “Usage is measured in data tokens,” input tokens covering context, instructions and history, output tokens the generated descriptions, relationships and queries returned. Data Agents pricing sets the rates at “$3 per million input data tokens and $20 per million output data tokens” and offers “a free trial of all Data Cloud agents and advanced AI features through September 30, 2026.” Service accounts get no free per-user quota and are metered the same way a person is. Google publishes both pages today, and neither points at the other. Size this product off the pricing page alone and you miss a line item that starts charging in about six weeks.
| What the instruments count | What no instrument reports |
|---|---|
| Entries, glossary terms, data products, assets per product, aspects per entry | Did the agent find what it needed |
| DCU-hours burned, gibibyte-hours stored, and from October 27, 2026, data tokens spent generating descriptions | If it did not, why not |
| Per-minute quota headroom on reads, writes, search and scans | Whether the failure was the model, the context or the harness |
Knowledge Catalog is a serious move on the left column. State the negative finding precisely, though: no answer-quality or retrieval-success meter is published in any documentation we could find, checked 2026-09-16 against the quotas page, the pricing page, the FAQ’s data insights billing section, the AI agents overview and the data insights page, which carries no release stage of its own either. Google is not standing still: verified queries and semantic guardrails, labeled Preview on the launch post, pair verified SQL patterns with pre-generated natural language questions against hallucinated logic and guessed joins, and the same post names “measurable context evaluation” with no status label and no documentation page behind it we could find. If that ships as an answer-quality metric, the gap narrows; until it does, a coverage dashboard cannot say whether a wrong answer came from missing context, poor retrieval or the wrong model, the attribution problem behind metric drift in enterprise text-to-SQL, inherited by the whole category.
For a single-vendor estate the bundled answer is defensible. A usage meter and a platform subscription still price different things, the subject of context layer TCO across build, buy and bundle. The cost on neither invoice is re-deriving your definitions in the next platform, what single-stack lock-in against a neutral context layer measures.
What a context layer adds above a cloud-native catalog
A context layer holds the decisions a query engine cannot observe, served to any agent through one connection.
Four parts form a loop. Context Agents create the context from evidence. The Context Lakehouse stores and vectorizes it in open tables in your own cloud. MCP and conversational AI serve it, one connection for agents and conversation for people. A learning loop finds what was missing when a question went unanswered and routes the fix back to the Context Agents, a sequence with no counterpart in a coverage meter, laid out in a context layer reference architecture.
Search alone does not get there. The vectorized layer serves vector, semantic, hybrid and graph-traversal retrieval over MCP, SQL and APIs, and graph traversal matters most, because “if I change this column, what breaks” is a path problem, not a ranking problem, where a context layer and a knowledge graph come apart. Most enterprises already run several definitions of the same metric across a warehouse, a BI tool and a notebook, each behaving locally as the authoritative one, which is why telling a context layer, a data catalog and a semantic layer apart comes first.
Atlan connects to Knowledge Catalog directly: it crawls aspect metadata, ingests quality and profiling results at column level, and writes aspect field values edited in Atlan back to the entries they came from. Google’s own launch post lists third-party business data catalogs, Atlan among them, as partner integrations. What the layer adds is the half Google Cloud never observed: the definition agreed in a review, the approval in a ticket, the owner accountable when the number is wrong. Atlan AI Labs found that agents grounded in governance metadata achieve 38% higher SQL accuracy than agents working from raw schema alone, the instrument shape every coverage meter lacks. Agree the definitions and make them portable first, then decide what serves them, the order in how to implement an enterprise context layer for AI.
Real stories from real customers: context an agent can reach
Both stories below are about a shared definition an agent can resolve without asking a human first.
53 % less engineering workload and 20 % higher data-user satisfaction
"Kiwi.com has transformed its data governance by consolidating thousands of data assets into 58 discoverable data products using Atlan. 'Atlan reduced our central engineering workload by 53 % and improved data user satisfaction by 20 %,' Kiwi.com shared. Atlan's intuitive interface streamlines access to essential information like ownership, contracts, and data quality issues, driving efficient governance across teams."
Data Team
Kiwi.com
🎧 Listen to podcast: How Kiwi.com Unified Its Stack with Atlan
One trusted home for every KPI and dashboard
"Contentsquare relies on Atlan to power its data governance and support Business Intelligence efforts. Otavio Leite Bastos, Global Data Governance Lead, explained, 'Atlan is the home for every KPI and dashboard, making data simple and trustworthy.' With Atlan's integration with Monte Carlo, Contentsquare has improved data quality communication across stakeholders, ensuring effective governance across their entire data estate."
Otavio Leite Bastos, Global Data Governance Lead
Contentsquare
🎧 Listen to podcast: Contentsquare's Data Renaissance with Atlan
What you actually own, as of September 2026
Four names, one endpoint, a migration some estates still owe, published limits that shape your design, and two meters that count supply.
The renames track real work. Automatic ingestion across Google Cloud’s own sources, extraction that writes descriptions without anyone typing them, an agent-readable surface with a dedicated lineage endpoint: a catalog was always supposed to do these things and mostly could not. What is missing is one instrument. Every published meter counts how much context exists, and none reports whether an agent asking a question left with the right answer. Until something does, coverage stands in for correctness, and coverage has never been a good proxy. Closing that is what an enterprise context layer is for. The next instrument Google ships will either count more context or report whether the agent found it. Worth watching which.
FAQs about Google Knowledge Catalog
1. Is Knowledge Catalog the same thing as Dataplex?
Yes. Dataplex Universal Catalog was renamed Knowledge Catalog on April 10, 2026, and Google’s product page is still titled “Knowledge Catalog (formerly Dataplex).” The pricing page uses the same construction, and its SKU names still read Dataplex.
2. Why do the API, CLI and documentation still say “dataplex”?
Because the rename did not change them: Google’s release note states the API, client library, CLI and IAM names remain unchanged. The endpoint is still dataplex.googleapis.com, the CLI still gcloud dataplex, the roles still roles/dataplex.catalogAdmin, and the docs still sit under /dataplex/docs/.
3. Do you still have to migrate off legacy Data Catalog?
If you hold no custom metadata, tags, tag templates, or custom entries and entry groups, setting the default catalog UI to Knowledge Catalog completes the move. Otherwise, a two-phase upgrade is on you. Data Catalog began a phased shutdown on June 1, 2026, and the transition guide was still live on September 15, 2026.
4. What are the limits on glossaries and data products in Knowledge Catalog?
Google publishes 5,000 terms and 200 categories per business glossary, nesting capped at three levels; 200 data products per project per region with 50 assets each; 10,000 aspects and 5 MB per entry; and search capped at 500 results in keyword-only mode or 100 otherwise.
5. What changes about Knowledge Catalog pricing on October 27, 2026?
Active billing for data insights starts, measured in data tokens at $3 per million input and $20 per million output, after the free trial of Data Cloud agents and advanced AI features ends September 30, 2026. Service accounts get no free per-user quota; processing, storage and the no-charge API calls are unchanged.
Sources
- Knowledge Catalog overview, Google Cloud
- Knowledge Catalog FAQ, Google Cloud
- Knowledge Catalog release notes, Google Cloud
- Knowledge Catalog deprecations, Google Cloud
- Transition from Data Catalog to Knowledge Catalog, Google Cloud
- Transition to business glossary on Knowledge Catalog, Google Cloud
- Data Catalog deprecation banner, Google Cloud
- About metadata management, Google Cloud
- Quotas and limits, Google Cloud
- Best practices for Knowledge Catalog, Google Cloud
- About data products, Google Cloud
- About MCP servers in Knowledge Catalog, Google Cloud
- Use the Knowledge Catalog remote MCP server, Google Cloud
- Knowledge Catalog MCP reference, Google Cloud
- Use the data lineage remote MCP server, Google Cloud
- Knowledge Catalog for AI agents, Google Cloud
- About data insights for structured data, Google Cloud
- Knowledge Catalog pricing, Google Cloud
- Data Agents pricing, Google Cloud
- Knowledge Catalog product page, Google Cloud
- Introducing the Google Cloud Knowledge Catalog, Google Cloud
- Data catalog versus context layer, Atlan
