LoginSign Up
← Back to Analytics
Weekly · Analytics · Week of August 3, 2026

Analytics rebuilds around agents: dbt points Fusion at lakehouse catalogs while MCP endpoints spread

mcpai-agentslakehouse-catalogsgovernancebi-as-code
Y
Generated 1h agoDrawn from 9 products

The week in analytics

The strongest signal in analytics this week was dbt Labs shipping on two tracks at once. dbt Core posted the sector's highest velocity by graduating its Fusion 2.0 alpha into read-write Iceberg REST catalogs — Fusion can now write to Horizon and Unity catalogs over Iceberg REST via DuckDB, and catalogs.yml v2 makes DuckLake, Iceberg REST, and local filesystem first-class catalog types. On the Python line, 1.12.0 went the other way, reverting the experimental dbt login command and the bundled dbt-state plugin rather than promoting them. The division of labor is now explicit: 1.x consolidates and removes, while 2.0 is where new surface area lands, and that surface is the open table catalog.

Underneath dbt, one pattern repeated across nearly every mover: analytics tools are being rebuilt so an agent can drive them, and the governance to trust that agent arrives in the same window. Lightdash and Neo4j matched dbt's velocity with two sparks each, Neo4j putting a hosted Model Context Protocol endpoint inside Aura and Lightdash letting its data apps reach external APIs through a server-side proxy. The MCP thread ran through OpenMetadata and MotherDuck too, while agent-as-interface work surfaced in Hex, Basedash, and — from a standing start — Sigma Computing.

Leaders

dbt Core led on both count and consequence. The Fusion 2.0-alpha.4 spark upgrades an earlier read-only integration into a write path — the difference between reading a lakehouse and building into one — and alpha.5 followed with Redshift datasharing catalogs and job-specific deferral. The 1.12.0 spark, by contrast, is a subtraction: two experiments pulled rather than shipped, which reads as the Python line clearing ground before 2.0 leaves alpha.

Lightdash shipped close to daily and split cleanly in two. Data App Connections let apps call third-party HTTP APIs without ever handling a credential, and a second spark lets you describe a chart type in a prompt and have Lightdash generate, test, and register it as a first-class visualization. Six improvements — content-as-code exports, a query inspector, timezone fixes — build the review surface underneath, aiming both humans and agents at one trust layer.

Neo4j is repositioning Cypher from the interface to an implementation detail. MCP for Aura is a hosted service authenticated with existing credentials, giving clients schema, read, and read-write tools with every response coming from a real query. A Document Intelligence preview builds the graph from unstructured files in S3, Azure Blob, or Google storage, and Cypher Copilot now lints and EXPLAIN-retries its own generated queries.

OpenMetadata promoted MCP to a first-class service category in 1.13.0 — service and server entities, execution logs, test-connection support, and usage analytics down to per-user tool breakdowns — while an RDF knowledge graph gave the catalog a standard query surface. The rest of the window is 1.12.x and 1.13.x maintenance plus a 2.0.0 release candidate cut for dev and test only.

Hex pulled its agent out of the notebook. The Hex Agent is now callable from a CLI and an API, with cell outputs in the terminal and a list of every resource a run touched — an audit trail that the prior two months of governance work were building toward. Claude Opus 5 joined a deliberately swappable model picker alongside generative apps as the artifact the agent produces.

Wildcards

Sigma Computing registered zero velocity but two sparks, the sharpest being a plugin for Claude Code that builds complete data models — metrics, relationships, columns, descriptions — from the terminal without opening Sigma. Paired with Automated Actions, which run analytics and AI workflows on a cadence, both sparks point at Sigma running without someone watching it.

Airbyte also showed no velocity but landed the window's most consequential release: 2.0 took data activation to general availability, moving modeled data out of the warehouse and into CRMs, marketing platforms, and support tools, with syncs four to six times faster. After a 1.x line spent making ingestion dependable, 2.0 runs the same pipeline in reverse.

Themes that compounded

  • MCP as a governed asset, not a bolt-on. Neo4j hosts it inside Aura, OpenMetadata made it a first-class service category with usage analytics, and MotherDuck feeds shared Guides into every agent session.
  • The agent moves outside its own UI. Hex exposed a CLI and API, Sigma Computing authors models from Claude Code, and Basedash put its entire feature surface behind a public API.
  • Lakehouse catalogs as the new integration target. dbt Core points Fusion at Iceberg REST catalogs directly, and MotherDuck added server-side Iceberg attach and R2 read/write.
  • Governance shipping half a step behind reach. Audit logs, SCIM, per-instance SSO, and approval prompts landed across Lightdash, Basedash, and Neo4j in the same windows as the capabilities they cover.
  • Reverse and outbound data flow. Airbyte 2.0 activation and Lightdash API-calling data apps both push analytics outward into operational tools rather than pulling data in.

Watch this week

Watch whether the alpha and release-candidate work in flight this window converges. dbt Core is still expanding catalog coverage adapter by adapter, so more catalog integrations and catalogs.yml v2 work should land before Fusion 2.0 leaves alpha, and OpenMetadata's 2.0.0 candidate should move through further RCs while 1.13.x keeps absorbing connector and CVE fixes. On the agent side, the open question is permissioning: Neo4j's read-write MCP access and Basedash's agent Actions both now record writes but describe approval case by case, so per-action policy is the natural next step to look for.