Direct SaaS MCP servers vs. a managed data layer for AI agents
Choose the right path to business context: direct application tools for actions, or Combined for SQL over synced, permissioned business records.
Choose Combined when your agent needs a managed place to query synced business data through MCP. Use a direct application integration when the job calls for that application’s own operations. MCP is the connection protocol in either case; the useful difference is where the records live, what the agent can do, and who keeps the data ready.
Start with the business question
“Create a follow-up task for this deal” and “How much open pipeline do we have by stage?” need different capabilities. The first needs an application action. The second needs the relevant records, a clear definition of open pipeline, and a query. Pick the data path that makes your recurring question easy to answer.
| Your agent’s job | A useful starting point | What to check |
|---|---|---|
| Create a task or update a CRM record | The application’s direct integration | The available write tool, authorization scope, and confirmation flow. |
| Filter, aggregate, or join managed business records | Combined’s data layer and read-only SQL through MCP | Source coverage, logical schemas, grants, and the last successful sync. |
| Find passages with similar meaning | A semantic retrieval system | Retrieval coverage and how the answer cites the original material. |
Direct MCP servers vary. For example, HubSpot documents both reads and writes through its remote server. A direct integration can answer useful questions inside that application. Combined becomes compelling when you want managed records, a consistent SQL interface, and scoped access for the agent across your data workflow.
Give the agent one path to managed business data
Combined brings source connection, sync, managed storage, and agent access into one workflow. Choose a Ready connector, authorize the business account, and select the datasets and fields you need. After the first sync, check the records in Explore and grant the agent access to the relevant data.
The agent can then discover sources and schemas, inspect freshness, and run bounded, read-only SQL over granted logical relations. Use the discovered relation names and fields instead of hard-coding another workspace’s table names. Add a second source when the question needs it, then define the join key and confirm both datasets are available to the same agent.
This is useful for developers building a business assistant, forward deployed engineers connecting client systems, and teams giving coding agents company context. You keep your agent and its interface. Combined supplies the managed data path behind the answer.
The agent query surface is read-only. Keep application-changing actions in an appropriate write integration. Combined’s search_context tool provides literal text lookup; use a semantic retrieval system when similarity search is the requirement. See SQL and vector search for business data for the distinction.
Make the first answer concrete
Start with open pipeline. In the HubSpot tutorial, four sample deals include two open opportunities worth $12,000 and $8,000, a $5,000 closed-won deal, and a $3,000 closed-lost deal. The open pipeline is $20,000 across two deals.
Ask the agent to discover the deals schema, exclude stages marked won or lost, group the open deals by stage, and aggregate their amounts. Keep currencies separate. If your custom stage IDs need a mapping, supply it before interpreting the total. A few sampled rows are useful for checking a connection; they are not the input for estimating the full pipeline.
Have the answer include the last successful sync time and query receipt. The result describes the records available at that sync. If someone has just changed a deal in HubSpot, sync that change before expecting it in the answer. Choose the available schedule that fits the job, using the freshness guide to set expectations.
Choose Combined for the questions you ask every week
If your team keeps exporting business records, rebuilding connector logic, or assembling the same context for every agent session, put a managed data layer behind the workflow. Combined lets you connect the source once, keep the selected data synced, and ask through the agent you already use.
- Choose one recurring question and the Ready connector that covers its data.
- Authorize the source, select the fields, and inspect the first sync.
- Grant the agent access and follow its account-specific MCP setup.
- Run the first query, then expand the data scope when the next question needs it.
Follow the Claude Code connection guide or the setup instructions in your Combined workspace. The initial 5 million MAR allowance does not require a card. Usage after the trial is $5 per million MAR.
Sources and further reading
Explore the documentation behind this guide. Product details checked on September 9, 2026.