For Databricks

Databricks governs the definition. You add the context. Your AI agent gets both.

ClariLayer imports your Databricks Metric Views as canonical definitions, then layers them with your own working context — so your agent recalls both in-flow and flags where your version has drifted from the governed one.

Connect your AIRead the Databricks quickstart
Claude CodeCursorCodex
ClariLayer with Databricks — where it fitsYour AI agent (Claude, Cursor, Codex) sits above your Databricks warehouse, connected to it by a dashed line meaning the agent runs live SQL with its own access and no shared credentials. A solid line over MCP connects the agent to ClariLayer, your context layer, which runs bootstrap, recall, remember and reconcile and shows a Diff to Canon panel where a local definition differs from the imported Databricks Metric View, flagged on recall. Status is caveat or asserted, never verified. Two arrows lead to outcomes: your agent grounded, and the Context Console.context · MCP, in-flowdata · your agent’s own accessin-flow contextMCPno credentialslive SQLown accessYour AI agentClaudeCursorCodexDatabricksUnity Catalog · Metric Viewsthe governed definitionClariLayeryour context layer · over MCPbootstraprecallrememberreconcileDiff to Canonlocal: arr_monthlycanon: committed_recurring_revenue⚠ filters differ — flagged on recallstatus: caveat or asserted · never “verified”Your agent, groundedstops citing the stale defwrong filter · grain · tableContext Consolecompare · adopt · variantstatus · provenance · where-used
Your agent talks to ClariLayer over MCP and to Databricks with its own access. The governed Metric View becomes canon inside ClariLayer; your local definition is checked against it.

The reconcile moment

The SQL is usually fine. The context is what breaks.

Your agent’s working answer

$1.51M

local arr_monthly note · last edited 6 weeks ago

Databricks Metric View

$1.42M

committed_recurring_revenue · governed

$90K apart — and without ClariLayer, the agent never flags it.
caveat recorded · filters differRecall surfaces the drift before you use the number. You adopt the Metric View; your agent stops citing the stale one.

How it works

Import once. Grounded from then on.

  1. 01

    Import your Databricks Metric Views

    In this personal MCP workflow, your agent reads the definitions from Unity Catalog with its own access and bootstraps them into ClariLayer as semantic_model canon — one entry per measure, canonical from the start. ClariLayer does not hold a source credential or touch your catalog directly on this path.

  2. 02

    Your agent recalls in-flow — no tab-switching

    Ask a data question inside Claude Code, Cursor, or Codex and your agent pulls your working context and the matching governed canon together, in one response. You never leave the editor.

  3. 03

    Drift surfaces before the number is wrong

    A different grain, base table, or filter shows up in the recall response, named: filters_differ. It works even when your local definition has a different name — ClariLayer matches on what a metric computes (table, measure, aggregation), not its label.

  4. 04

    Compare, adjudicate, reconcile

    Open the console: adopt the governed definition (your SQL is preserved), link it as a deliberate variant, or keep local. Then reconcile against your warehouse for an honest asserted or caveat outcome — never a verified badge.

Compare with canonical

See exactly where your context drifts

Your local def arr_monthlyMetric View committed_recurring_revenue
base tableanalytics.bookingsanalytics.bookings
measureamountamount
aggregationSUMSUM
grainmonthmonth
filters— none —exclude in-quarter discounts
filters_differMatched on what they compute, not their names.

Privacy & access

Personal MCP access stays with your agent

In this personal workflow, your agent is the connector. It keeps its own Databricks access and sends definitions and supported result evidence — metadata plus any optional preview rows it chooses to include. ClariLayer receives no source credential and runs no server-side query on this path.

When you import, your agent passes the structured definition; when you reconcile, it runs the SQL and sends back the result shape. On this personal path, no warehouse credentials reach us and no SQL runs on our servers. The gated Governed Context Edge can use a team connector only when the participating team explicitly configures it under that surface’s own controls. The full data-flow posture is on the security page.

Honest scope

What the personal MCP path doesn’t do

Personal path: no direct warehouse access

In the personal MCP workflow, ClariLayer has no Databricks connection of its own. It does not read Unity Catalog, query Delta tables, or ingest query history; your agent does that with its own access.

Personal path: no server-side queries

When you reconcile a metric in the personal MCP workflow, your agent executes the SQL and sends back the result shape. ClariLayer does not initiate a query, schedule a scan, or hold a source credential on this path.

Only marks it can back

Context entries are asserted or caveat: a mismatch is flagged as a caveat, and the stronger “verified” mark is not live or promised on a delivery timeline. Every mark you see today is one we can back.

Who it’s for

Built for the individual Databricks analyst

You work a Databricks workspace in Claude Code, Cursor, or Codex, with Metric Views in Unity Catalog. Single-player — no data team, no procurement, no shared account. Install the MCP server, run one bootstrap, and your agent is grounded on your canon.

Ground your agent on your Databricks canon.

Connect ClariLayer to Claude Code, Cursor, or Codex. Import your Metric Views, and your agent stops guessing — it recalls the governed definition and flags your drift in-flow.

Connect your AIRead the Databricks quickstart