Book a demo Sign up
Kroolo Search · Embedded Analytics

Embedded analytics, answered from your own data.

Put governed dashboards and an analyst that answers in plain language inside your product. Kroolo queries your warehouse in place, as the viewer, and shows the SQL, lineage and freshness behind every number.

Your product · Portfolio insightsExportShare
Written premium
$48.2M
+6.1% QoQ
Claims ratio
61.4%
-1.8 pts
Renewals at risk
37
+9 this week
Data as of
06:12
dbt run passed
Written premium by quarterEntitlement-aware · cached
Why did renewals at risk rise this week?
Mostly the SME segment in the north region. Nine policies lapsed after the pricing change on Monday. View SQL · Lineage · Metric definition
Query in place
no bulk copy of your data
Per-user
identity from question to source
4 tiers
of agent authority, with approvals
Open source
components only
Your data platform

Connect what you already run.

Warehouses and lakehouses are queried where they sit. Pipeline tools feed lineage, freshness and definitions, so answers come with context.

Snowflake
Databricks
Amazon Redshift
Microsoft Fabric
dbt
Airflow
Architecture

Built in layers, governed at one gate.

Requests flow from the experience down through agents and a single tool gateway into context and query layers. Identity, secrets, observability and audit are shared by all of them.

Experience
Embedded dashboardsAsk-anything analystAPI and MCP clients
Agents
AnalystDashboard builderMonitorPipeline operatorCurator
Gateway
Authentication, policy checks, limits, approvals and audit on every tool call
Context
Catalog graphRetrieval indexMetric layerFreshness and trust
Query plane
Query routerNative adaptersFederationResult cache
Sources
SnowflakeDatabricksRedshiftMicrosoft FabricLakesdbt · Airflow
IdentitySecretsObservabilityAudit store
Context

Teach it your business once.

A connected source is not useful until the agent knows what the tables mean. Kroolo builds that context from your catalog, dbt and pipelines, and keeps people in charge of definitions.

Catalog graph of tables, columns, joins, owners and lineage
Governed metrics through MetricFlow, so “written premium” always means the same thing
Freshness and trust signals from dbt and pipeline runs
Agents draft descriptions, a person approves them
Governed context
written_premiumSUM(written_premium) · metricGoverned
loss_ratioclaims ÷ earned premium · metricGoverned
regiondimension · 5 valuesDimension
fct_premiumlast dbt run 06:12 · tests passedFresh
stg_claims_rawdeprecated · owner: data platformAvoid
How this number was produced
Metric resolved
written_premium from the governed metric layer
Policy checked
Viewer may read fct_premium
Ran in Snowflake
As the viewer, after a dry-run cost check
Profiled
Q3 drop in one region flagged
Cached
Keyed to the viewer’s entitlements, refreshed daily
Trust

Every number shows its work.

Answers link to the SQL, lineage, freshness and metric definition behind them, so a finance lead or an auditor can follow the figure back to source.

SQL, lineage and data-as-of time on every answer
Anomalies flagged with simple, explainable statistics first
A locked state, not an error, where a viewer lacks access
Exports carry the data-as-of timestamp and dashboard version
Dashboards

Dashboards are specs, not code.

Agents describe a dashboard as a versioned document. That makes it easy to validate, diff, roll back and refresh, and it stays tidy as people keep asking for changes.

Agents edit the spec, not UI code
Results cached as Parquet and cross-filtered in the browser
Refresh on a schedule and when the upstream pipeline finishes
Public links are off by default
dashboard-spec.json
{
  "dashboard": "portfolio-insights",
  "version": 14,
  "datasets": [
    { "id": "premium", "metric": "written_premium",
      "dims": ["region", "quarter"], "refresh": "after:dbt_premium" }
  ],
  "charts": [
    { "type": "bar", "dataset": "premium",
      "x": "quarter", "y": "written_premium", "color": "region" }
  ],
  "layout": [ ["premium_bar", "kpis"] ],
  "access": { "locked_state": true, "public_link": false }
}
Permission tiers
1 · ReadCatalog search, read-only SQL, dashboardsNo approval
2 · MonitorScheduled queries and alertsApprove once
3 · TriggerRun allowlisted dbt or Airflow jobsEvery run
4 · Write-backScratch tables or a dbt pull requestReviewed PR
Actions

Agents that act, with your approval.

Reading, monitoring, triggering and writing back are separate permission tiers, each with the controls its risk deserves. Anything that changes state outside Kroolo needs approval.

Dry-run first: tier 3 actions preview before they run
Every proposal is an action card with parameters, as whom, and the cost
Write-back goes to a scratch schema or a reviewed pull request
Kill switch and rate limits for your administrators
Embed it

Three ways to put it inside your product.

One permission model sits behind all three, so there is no second set of rules to maintain.

Embedded dashboards
Put governed dashboards inside your own product, with each viewer seeing only what their entitlements allow.
Pipeline created, by monthThis year
Aug
Sep
Oct
Nov
An analyst in your product
Let customers and colleagues ask the next question in plain language, with the same metric definitions as the dashboard.
Summarise open risks this week
Kroolo AI
2 risks are high: vendor delay and budget.
API and MCP server
Give your own assistants and agents governed access to the same tools, with the same policy checks and audit.
GR1Auth context enforcedCritical
GR2Source citation requiredHigh
GR3PII masked in outputMedium
Security

Built for financial-services scrutiny.

Controls designed with banks, insurers and regulators in mind. Compliance mappings are an engineering aid, not a certification.

Per-user identity
The end user’s identity reaches the source wherever possible, so source permissions, row filters and masks apply.
Entitlement-keyed cache
Cached results are keyed to what the viewer may see, and encrypted with per-tenant keys.
Data minimisation for models
Sensitive columns are never sent raw, full result sets stay out of prompts, and open models can run in your VPC.
Immutable audit
Who saw what, through which agent, when, and why it was allowed, exportable to your SIEM.
Prompt-injection defences
Data from sources is treated as untrusted. Authority sits in the gateway, not the model.
Residency and retention
A control-plane region per tenant, configurable retention and deletion on offboarding. SaaS first, with a customer-VPC edition designed in.
Why Embedded Analytics

Fast to ship, safe to trust.

Ask in plain language
Viewers ask the next question and get an answer, not a ticket to the data team.
Summarise open risks this week
Kroolo AI
2 risks are high: vendor delay and budget.
Dashboards from a spec
Agents build and edit dashboards as versioned specs you can review and roll back.
Pipeline created, by monthThis year
Aug
Sep
Oct
Nov
Per-user access
Every query carries the viewer’s identity, so source permissions and masks apply.
Access, enforced at retrieval
Finance teamRead
Sales leadsRead, edit
ContractorsNo access
Your definitions
Governed metrics through MetricFlow keep every answer consistent.
Which deals will slip this quarter?
Kroolo AI
4 deals ($84k) have no activity in 14 days.
Fresh by design
Refresh on a schedule or when the upstream pipeline finishes, with data-as-of shown.
This weekScheduled
Mon
Tue
Digest
Wed
Thu
Report
Fri
Auditable
Who saw what, through which agent, and why it was allowed, all recorded.
Access, enforced at retrieval
Finance teamRead
Sales leadsRead, edit
ContractorsNo access

FAQs on Embedded Analytics

Does Kroolo copy our data?
No. Data stays in your platform and is queried in place. Results are cached, not replicated, and cache entries expire and are keyed to what each viewer may see.
Which platforms does it connect to?
Snowflake, Databricks, Amazon Redshift and Microsoft Fabric, with lakes on S3, ADLS and GCS. dbt and Airflow feed lineage, freshness and definitions.
How do viewers only see their own data?
The end user’s identity reaches the source wherever possible, so row filters and column masks apply. Cached results are keyed to entitlements, and a locked state shows where access is missing.
Can agents change anything in our systems?
Only through approved tiers. Reading and monitoring are low risk, and triggering a pipeline or writing back needs approval, with a dry run, a kill switch and a reviewed pull request for write-back.
Can we run it in our own VPC?
The product starts as SaaS, with the control and data planes separated so a customer-VPC edition is designed in.
How is accuracy measured?
Each customer environment has golden questions with reviewed expected answers. Changes to prompts, context, models and routing must pass those gates before release.

Ship analytics with Kroolo Search.

See embedded analytics on your own data in a demo.

Book a demoExplore Kroolo Search