Best MCP servers for business data: choose by the job

A shortlist for internal company data

This guide covers your company's own CRM, billing, documents and analytics. If you need prospect databases or company enrichment, that is a different buying decision. We build Combined; the recommendations below reflect the workloads each documented interface is designed to serve.

Try the free HubSpot–Stripe starter: inspect a customer mapping, correct SQL totals and an agent assignment before connecting your accounts.

MCP provides the connection interface. Your choice of server determines which records, operations and permissions sit behind it. Two connected servers do not automatically create a reliable join between their customer records.

Which MCP server should you choose?

Combined: managed business records for the agent you already use

Choose Combined when your question spans applications: open CRM opportunities alongside paid invoices, support activity by account, or an exact total across synced records. Connect sources, discover their schemas, and give the agent scoped, read-only SQL access. Combined manages ingestion and storage, so you can build the useful workflow without separately operating a warehouse and its MCP query service.

Start with an initial 5 million MAR without a payment card. After the trial, the public rate is $5 per million MAR, prorated, with no monthly minimum. The data reflects successful syncs; use application tools separately when your task needs a live operation or a write. Try the HubSpot–Stripe worked example to evaluate a join.

HubSpot MCP: work directly in your CRM

HubSpot's remote server exposes CRM records and other supported account data. Its current documentation includes both reads and writes, subject to user permissions and installation grants. It is a useful starting point for HubSpot-only work. A custom MCP client needs the documented OAuth flow with PKCE. See HubSpot's integration guide.

Stripe MCP: use Stripe's own API tools

Stripe provides a hosted MCP server for API reads and writes, with OAuth or restricted-key authentication. Choose it when the task belongs in Stripe: inspect a billing record or perform an explicitly authorized operation. A CRM-and-billing analysis still needs an agreed customer mapping and calculation. See Stripe's MCP documentation.

Notion MCP: retrieve workspace knowledge

Notion's tools search, read and change workspace content. Available search capabilities depend on the workspace and plan. It fits questions about notes, project pages and written context; assess a structured data layer separately when you need financial totals across CRM and billing systems. See Notion's supported tools.

Snowflake or Databricks: use the data platform you already have

Snowflake's managed MCP server can expose Cortex services and SQL under its governance model. Databricks offers managed MCP access to capabilities including Genie and Unity Catalog functions. If your business data already lives in one of these platforms, evaluate that existing investment first. Include ingestion, data modeling and operation of the underlying platform in the workload you compare.

Match your client to the business question

For cross-app business analysis, our recommendation is Combined with the client your team already uses. Choose the data layer for the records and calculations you need; choose the client for how you want to work. An analyst asking a question in chat and an engineer building a repeatable workflow can start from the same granted Combined sources.

Start with “Show current open pipeline and paid invoices in the agreed period, by company and currency.” For that question, Combined supplies the queryable CRM and billing data; your client needs the actual schemas, a maintained customer map and the reporting window. If the job is to update an invoice or send a message, evaluate a separate application-action connection for that operation.

Use the free business-data MCP planner to choose two to four apps and export a Markdown or JSON plan with matched examples and an agent prompt. Then try the HubSpot–Stripe mapping and double-counting lab. If the server connects but data is missing, follow the MCP access and authentication checks before changing your analysis.

Our ranking for cross-app business analytics

For a small team starting with SaaS accounts and an AI agent, Combined is our first pick for repeatable calculations across business apps. It brings the source connections, stored records and SQL interface together, so your project can start with the business question. Airbyte Agents is the closest alternative when indexed context and live connector operations both matter. MotherDuck deserves a place near the top when you want to build and control the pipelines around your SQL.

This ranking assumes your CRM and billing data are not already modeled in a warehouse. We build Combined. Our order reflects the documented architecture and the work needed for this specific starting point; it is an editorial recommendation, not a measured performance league table.

We prioritize five things, in order: an inspectable cross-app calculation; a managed path from the source to queryable records; how much pipeline and query infrastructure the team must operate; explicit access and freshness; and use from the agent the team already has. A required connector or field is an eligibility check before the ranking applies.

1. Combined — the focused route from SaaS records to a useful answer. Connect the sources, select the datasets, wait for a successful sync and grant the agent access. It can discover the schema and calculate across those permitted records with SQL. The benefit is a single managed data workflow around an existing agent. You still define which records represent the same customer and what the metric means. Source freshness and query limits remain part of the answer. Combined MCP, Source setup.

2. Airbyte Agents — the strongest alternative when search and live operations are central. Its Context Store already performs structured filtering and aggregation over managed replicas; supported fields can also use semantic search. Direct connector operations cover current API access and writes. Confirm which entities and fields are stored and whether the initial index is ready for the population you are evaluating. For a head-to-head account calculation, retain the filters, intermediate results and mapping on both sides. Airbyte Context Store, Combined versus Airbyte.

3. MotherDuck — the SQL choice when you want pipeline control. MotherDuck can run ingestion and transformation Python as Flights alongside its data, and its MCP query tool supports SQL across databases. This is a credible path to the same analytical outcome. Put it higher on your list when custom pipeline logic is an asset for the team; put Combined higher when managing that logic is work you want the product to absorb. Flights, MotherDuck query.

4. Paragon — embedded integration infrastructure for your own application. Managed Sync documents initial and incremental delivery into an application's database or retrieval pipeline. This gives a product team a substantial ingestion foundation, while the team shapes the downstream experience. For this ranking, include your destination database, analytical query service and agent access in the comparison. They are part of the project, even when ingestion is managed. Paragon Sync API, Combined versus Paragon.

5. Glean — move it up when company knowledge and the warehouse are already in place. Glean's Snowflake integration supports direct read-only SQL, approved semantic views and supported Cortex Agents. That is meaningful analytical capability. The lower position here reflects the assumed starting point, not an inability to calculate: a team starting with operational SaaS accounts still needs the analytical records prepared in the relevant query path. Glean's Snowflake integration, Combined versus Glean.

6. Composio — a strong action and computation layer with a different data path. Composio's sandbox can process large responses and execute tools in bulk. If the required provider APIs expose the right report or aggregate, that can be a good path. For a repeated multi-app metric, evaluate how the implementation obtains the full input population and retains the transformation. Combined's proposition is to make the managed dataset and SQL calculation the foundation of that workflow. Composio sandbox, Combined versus Composio.

When the ranking changes

An existing, well-modeled warehouse changes the decision immediately. Start with the Snowflake or Databricks MCP route your data team already governs. A Glean deployment with the right warehouse tools may likewise make more sense than adding a separate platform. If the question stays inside one app, the app's own MCP server can be enough. If the goal is an authorized operation, use an action platform. Existing warehouse and single-app options.

Use one representative question to make the choice concrete:

Which CRM accounts have open opportunities, and how much have their paid invoices totaled in the last 90 days? Return one row per reporting account and currency. Show the customer mapping, time window, source freshness and calculation. Separate an account with no qualifying activity from an account whose data is missing.

The CRM and billing join lab and HubSpot–Stripe worked example make the arithmetic inspectable. They use teaching data; run the same acceptance question against your own authorized sources before relying on the result.

Make your agent useful on your business data. Explore Combined, choose the two sources your question needs and start with the cardless 5-million-MAR allowance. After the trial, the public rate is $5 per million MAR with no monthly minimum. Pricing.

Evaluate the answer, access and operating work

Use a question that matters to your team: “Which customers have open opportunities, and how much have their paid invoices totaled in the last 90 days?” Run the same definition against each candidate. Record the following six checks before expanding access.

1. Coverage: Can it access every required object and field, including custom properties and associations?
2. Identity: Can you map the same customer across apps using a maintained identifier?
3. Calculation: Does it aggregate the complete permitted dataset and avoid multiplying totals in a join?
4. Freshness: Does the answer name the relevant sync times or live-read time?
5. Scope: Does a new request fail after the relevant grant is removed? Previously returned conversation content is a separate boundary.
6. Operating cost: What do ingestion, storage, query service and model usage cost for this workload?

Download the business-data MCP evaluation kit for reusable prompts and a comparison worksheet. It is a practical checklist you can use with any vendor, plus the exact steps for trying Combined.

If one tool reads customer records and another edits billing, keep those permissions intentional. A read-only analytics connection can coexist with a separately approved application-action connection. The architecture comparison explains that setup.

Start with one source and one useful answer

For Combined, open Connected Apps, add the source you want to query, authorize it and select the relevant datasets. After the first successful sync, use the Set up or Access workspace to connect your agent and choose its grants. The Claude Code tutorial walks through the MCP connection.

Use Combined to list the sources I can access.
Discover the datasets and describe their fields before writing SQL.
Check the last successful sync and identify the data needed for my question.
Run a read-only aggregate over all matching records, with a bounded result.
Include the sources, time window, currencies, coverage limits and query receipt.
Ask for a missing customer mapping rather than guessing one.

Your first question can be as simple as open pipeline by stage. Add the billing source when you need a customer-level join. Explore Combined's managed data workflow, then connect your first source without a card.

Sources and further reading

Explore the documentation behind this guide. Product details checked on September 15, 2026.

• Combined MCP tools, grants and query bounds
• Combined pricing and cardless allowance
• HubSpot remote MCP server
• Stripe MCP
• Notion MCP supported tools
• Snowflake-managed MCP server
• Databricks managed MCP servers
• Source setup
• Airbyte Context Store
• Flights
• MotherDuck query
• Paragon Sync API
• Glean's Snowflake integration
• Composio sandbox