Combined
GuidesDocsStart free trial
Combined
GuidesDocsStart free trial

Connect Claude Code to business data with Combined MCP

A practical setup guide: connect a source, grant access, add the MCP server, discover schemas and verify a real query receipt.

By the Combined team ·

To give Claude Code read-only access to your business data, first connect a Ready source in Combined and grant the required datasets. Then add Combined’s account-scoped HTTP MCP endpoint to Claude Code, authenticate, and verify an actual query. A connected server alone does not prove that the agent can read the right data.

1. Prepare one source and a narrow grant

You need Claude Code with HTTP MCP support, a Combined workspace, a Ready connector that covers your intended dataset, and permission to authorize that source. Start with a low-risk dataset appropriate for your account. Combined’s cardless allowance is 5 million MAR; check the source’s expected volume before its initial sync.

  1. In Combined, choose a Ready connector and complete authorization.
  2. Select the datasets and fields needed for the task, then run the first sync.
  3. Verify the data in Explore. Empty or incomplete data is a source problem to resolve before testing the agent.
  4. Use the onboarding Use it with AI step or the Access workspace to set up the agent’s identity and permitted datasets.

Use the product-generated setup instructions when available: they include your account’s endpoint. Keep the grant limited to the task. A human’s ability to see data in the workspace does not automatically grant the agent access.

2. Add the HTTP MCP server

Replace ACCOUNT_UUID with the account ID supplied by Combined. The account ID selects the account boundary; it is not an access credential.

claude mcp add --transport http combined-context \
  "https://platform.trycombined.com/mcp?account_id=ACCOUNT_UUID"

claude mcp list

In Claude Code, open /mcp and follow the server’s authentication flow. Combined advertises OAuth discovery through its protected-resource metadata. An initial HTTP 401 before login is part of that discovery flow, not by itself evidence that the service is down.

If your runtime needs a manually provisioned credential, generate the client-specific configuration from Access and provide the token through a secret environment variable. Never paste a live token into a repository, screenshot or shared setup prompt. See the MCP reference for the documented JSON configuration.

This walkthrough follows the published Combined and Claude Code documentation. Authentication details can vary by client version and workspace policy; the receipt check below is how you verify your own installation.

3. Discover the schema before asking a business question

Give Claude Code an instruction that makes the permitted scope explicit. This example deliberately avoids assuming any connector-specific table names:

Use the combined-context MCP server.
1. List the sources I am permitted to read.
2. Choose the source I specify and list its datasets.
3. Describe the selected dataset and report its freshness.
4. Run a read-only query returning at most 5 rows, using
   the logical relation and fields returned by discovery.
5. Report the receipt ID, freshness, and any truncation.
Do not guess table names or request broader permissions.

The tool sequence is list_sources, list_datasets, describe_dataset, then query_sql. Use get_freshness when the question depends on recency. Follow pagination when discovery returns a cursor.

Combined accepts bounded, read-only SQL over granted logical relations. Use parameters for values and identifiers from discovery. The query tool defaults to 100 rows and permits at most 1,000; lower maxRows for an initial check. A low output row limit does not remove input-scan or time limits.

4. Verify the connection with a receipt

Compare the agent’s small result with the corresponding permitted data in Combined. Check the account, source, selected fields and freshness. A successful response should include a real query receipt; an explanation that the agent “could query” the data is insufficient.

  • Server appears, no sources: inspect the agent identity, account ID and current grants.
  • Sources appear, expected rows missing: inspect the first sync, selected fields and source freshness.
  • Query rejected: check discovered relation names, permitted SQL and effective input limits.
  • Authorization fails after working: check whether access was revoked or the credential needs renewal. Do not widen the grant just to silence an error.

For a controlled revocation check, use a dedicated test grant and confirm that it stops permitting the query after revocation. Avoid interrupting an agent used by somebody else. See query and MCP troubleshooting for recovery.

Once this small test succeeds, ask a specific business question with a defined time window and data scope. Have the agent state missing coverage, stale sources and truncated results alongside its answer.

Sources and further reading

Explore the documentation behind this guide. Product details checked on September 9, 2026.