Connect your data stack

Connect the orchestrator, warehouse, transformation project, and code hosting that the RCA agent investigates with, by adding MCP servers to Claude Code.

The Soda plugin installs one MCP server: Soda MCP, which tells the agent which checks are failing. To find out why they fail, the agent needs to reach the rest of your stack. Each additional capability is an MCP server that you register with Claude Code once, at user scope, so that it applies to every session.

Capabilities

Capability	Question it answers	What the agent does with it
Orchestration	Whether the pipeline run actually did the work	Lists recent runs and reads task logs to catch a green run that skipped its work, and finds the deployed code version.
Warehouse	Whether the data is wrong, and where freshness breaks	Runs read-only queries for row counts and load timestamps, and reads the warehouse's own lineage where it has one.
Transformation	Which model produces the bad column	Reads the model graph and the model SQL, and finds the non-dataset consumers such as dashboards and reports.
Code history	Which change caused it	Reads the commits that touched the pipeline at the deployed version.

Code history is usually covered already: the gh and git CLIs on your PATH count as coverage, as long as the repository is reachable from your shell. On Windows, Git for Windows provides git.

How the agent reports gaps

At the start of every investigation, the agent checks which capabilities the session can reach. The check classifies each connected MCP server by the system named in the server's name, command, or URL. A missing capability never blocks the investigation. The agent works with what is present, states in the report which evidence it could not gather, and offers to help you close the gap: it asks which system you run, then suggests how to register its server.

The offer is made once per configuration. It is repeated when you add or remove a server, or after 30 days. If you tell the agent not to ask again, it records that and stays quiet until your set of servers changes.

Register an MCP server

Every server is registered the same way. Use --scope user so that the server is available in all your projects:

# A server that runs locally and talks over stdio
claude mcp add <name> --scope user -- <command>

# A server reached over HTTP
claude mcp add <name> --scope user --transport http <url>

Pass credentials in the server's own environment block with -e KEY=value, never on the command line of a shared terminal. The commands are the same in PowerShell on native Windows.

Name the server after the system it connects to, for example snowflake, airflow, or dbt. The agent's coverage check recognizes a system by its name in the server name, command, or URL. A server named prod-data that starts a Snowflake command is still recognized, but one that hides the system behind an unrelated name and a wrapper script is not.

For the full syntax, including project-scoped servers and OAuth, refer to the Claude Code MCP documentation.

After registering a server, run /mcp inside a Claude Code session to confirm it is connected. The next /rca investigation picks it up.

Where to find a server for your system

The tables below point to where each vendor documents its MCP server. Follow the vendor's current instructions for the exact command or endpoint. A system that is not listed here almost certainly has a server too: search the vendor's documentation for MCP.

Orchestration

System	Server
Airflow 3	The Airflow API server serves an MCP endpoint at /mcp/v1. Register it with --transport http http://<airflow-host>/mcp/v1.
Dagster	Refer to the Dagster documentation.
Prefect	prefect-mcp-server

Warehouse

Register warehouse servers read-only. Most servers accept a configuration that limits them to statements such as SELECT, SHOW, and DESCRIBE. An investigation must never be able to write.

System	Server
Snowflake	Snowflake-Labs mcp, run as uvx snowflake-labs-mcp --service-config-file <config>.yml with the configuration limited to Select, Show, Describe, and Use. It also exposes Snowflake's native lineage.
Databricks	Databricks managed MCP servers or Databricks Labs mcp
BigQuery, PostgreSQL, and others	MCP Toolbox for Databases
Redshift and other AWS services	AWS MCP servers

Transformation

System	Server
dbt	dbt-mcp, run as uvx dbt-mcp and pointed at your project. Its model graph also declares exposures: the dashboards, reports, and extracts that read your models.
SQLMesh	Refer to the vendor's documentation.

Without a server, a checkout of the transformation project on disk answers the same questions, more slowly. Tell the agent where the project lives when you start the investigation. It does not search your filesystem for one.

Code history

System	Server
GitHub	The gh CLI, or the GitHub MCP server
GitLab	GitLab MCP server
Bitbucket	Atlassian MCP server

A code hosting server is an alternative to the CLIs, not an upgrade. Add one only when the repository is not reachable from your shell.

To run an investigation with the connected servers, refer to Root cause analysis agent.

---

You are not logged in to Soda and are viewing the default public documentation. Learn more about Licensing & documentation access.

If you do have a Soda license, make sure to log in to Soda Cloud in this same browser.