Combined vs Databricks MCP: turn business apps into agent answers

Compare Combined with Databricks SQL, Genie, and AI Search MCP. Choose your data setup and evaluate paid customer retention across PostHog and Stripe.

By the Combined team · September 10, 2026

Choose Combined for managed sync and read-only SQL across business apps. Choose Databricks MCP when your team wants agents to use its existing Unity Catalog data, Genie definitions, search indexes, or platform functions. Both can support business questions; the useful choice is which data preparation and platform operations you want to own.

Start with the business question and the data you have

Your product team wants to know whether activated accounts remain paying customers. The activity is in PostHog; the payment evidence is in Stripe. Combined connects business sources, keeps selected records synced, and gives your agent managed SQL access. You can build that account-level answer without separately operating a lakehouse and its agent query service.

Documentation reviewed September 10, 2026 UTC; Databricks AWS documentation
Decision	Combined	Databricks MCP
Starting point	Business apps you want connected and queryable.	Workspace data and tools governed through Unity Catalog.
Analytical path	Agent schema discovery and read-only SQL over synced Sources.	Direct SQL, Genie analytics, AI Search, or registered functions.
Access model	Account scope and explicit Source grants.	Caller resource permissions and the selected endpoint or service.
Work to prepare	Connections, selected datasets, successful sync, and business definitions.	Data availability, compute, permissions, and the chosen analytical resources.
A strong reason to choose it	Manage the business-app-to-query workflow in one product.	Extend a Databricks platform and semantic layer you already use.

The table reflects the Combined MCP reference and Databricks' managed MCP overview. Existing data models and operational ownership matter more than counting endpoint names.

SQL, Genie, and search solve different parts of the job

Databricks SQL MCP is Public Preview in the reviewed AWS documentation. It executes permitted reads and writes against Unity Catalog tables, with asynchronous completion. Specify a warehouse or let the system select one from available resources and permissions. This is a direct SQL path for inspection, development, and explicit calculations. SQL MCP documentation.

Genie One is Beta and requires the workspace preview to be enabled. It accepts natural-language questions, uses Genie Ontology, writes SQL, and returns answers with source links. A single Genie Agent MCP instead scopes the question to a curated set of up to 25 tables; that interface is read-only and does not pass conversation history to the Genie API. Genie One; Genie Agent.

AI Search MCP provides semantic retrieval over a prepared index and requires Databricks-managed embeddings. It is Public Preview. Unity Catalog functions expose registered Python or SQL logic as tools. These are useful capabilities when your application needs document retrieval or reusable platform operations. AI Search; function tools.

Databricks also recommends the newer system.ai.dbsql MCP Service for additional governance. That SQL service is Beta; its policy can block writes. These release stages apply to specific paths, so check your target cloud and workspace. MCP Services; SQL service policy.

Choose how the records become ready for your agent

Databricks has managed SaaS ingestion through Lakeflow Connect, including providers such as HubSpot, Salesforce, and Zendesk. Verify the particular source, destination permissions, and compute prerequisites; the Salesforce pipeline, for example, requires Unity Catalog and serverless availability. SaaS connectors; pipeline prerequisites.

For a team starting from business apps, Combined brings connection, sync, storage, and query access into the same workflow. Check current Ready status and selected datasets in Connected Apps, complete the first sync, and grant Sources to the agent. Use list_sources and list_datasets with pagination, then describe_dataset. The live schema determines your query; example relation names are not guaranteed tables. Combined connection setup.

Combined's six read-only tools also include query_sql, get_freshness, and search_context. Query across the complete eligible records and display a bounded answer. Check successful sync coverage before treating missing data as zero. Your product-account map and paid-retention definition stay explicit, so the answer can be reviewed and reused. MCP tools and limits.

Measure paid retention with an account-level test

Open the PostHog–Stripe retention plan and download the platform evaluation worksheet. Ask which accounts activated in July, had paid service at August 1, and remained covered at September 1. Compare the same metric through Combined, a prepared Genie definition, or an explicit Databricks SQL query.

Resolve event actors to product accounts, then accounts to billing customers. Deduplicate repeated events and account/customer pairs. Freeze the denominator at accounts with paid coverage on August 1. New August customers must not inflate that cohort. Exclude staff, QA, and test billing records explicitly.

Paid coverage requires a paid invoice and positive payment attributed to recurring service. Use the relevant invoice line's service interval, with its start inclusive and end exclusive. An active subscription, an open invoice, or a paid setup fee beside a free recurring line does not establish this metric. These are the rules in our paid-retention workflow.

Build distinct product accounts with first_report_shared in
[2026-07-01, 2026-08-01), UTC. Validate the billing identity map.
At August 1 and September 1, require paid recurring coverage:
service_start <= checkpoint < service_end, with payment evidence.
Keep August baseline payers as the fixed denominator.
Return activated accounts, baseline payers, retained accounts,
retention rate, exclusions, source freshness, and SQL. Read only.

Run the synthetic DuckDB fixture against its expected result: three activated accounts, two baseline payers, and one retained payer, giving 50% retention. Birch's service ends exactly at September 1, so it is outside that checkpoint. Atlas's two users remain one account.

Adapt the fixture's teaching relations and dialect to actual datasets before either platform evaluation. Record the ingestion route, query, answer, and unresolved coverage. The fixture uses available paid records; a historical “what was known then” report also needs payment timestamps and identity history. These are synthetic expected results, not observed Databricks or customer outcomes.

Questions to settle before choosing

Do we need manual OAuth application setup for Databricks?

The documented manual OAuth route uses an account-admin-created application. The Unity Gateway CLI route can reuse an authenticated Databricks CLI session. Confirm resource permissions and network access for your client. Client setup options.

Is Combined's text search equivalent to AI Search?

Combined provides bounded literal substring lookup. Choose the appropriate indexed semantic-search route when that is your main workload. For retention arithmetic, focus the evaluation on account identity, paid service periods, and complete inputs.

Which is cheaper for this workflow?

Compare actual records, ingestion, storage, compute, optional search, and setup effort. Databricks managed MCP uses its underlying feature's pricing; Combined uses monthly active records. Existing platform work changes the incremental cost. Databricks pricing inputs; Combined pricing.

Give your product team a repeatable answer across activity and billing. Explore Combined, start with 5 million MAR without a card, and follow the self-serve setup. Use the retention plan to connect the right records and make the first calculation count.

Sources and further reading

Explore the documentation behind this guide. Product details checked on September 10, 2026.