Windows Preview available. Pro checkout is not open yet. Plans and capability details →
RowCoveAccount ↗

MCP tutorial

Connect deliberately.
Inspect every step.

A local tool server can save time when its access and outputs are explicit. Start with reads and a narrow folder.

1. Set up the server

  1. Activate Pro in the desktop app.
  2. Open MCP settings and enable the server. Limit the authorized roots to a folder containing the intended data.
  3. Copy RowCove's generated client configuration. It contains the actual executable and configuration paths for your installation.
  4. Add it to your MCP client's server configuration. Restart or reconnect that client as its own documentation requires.
{
  "mcpServers": {
    "rowcove": {
      "command": "<executable copied from RowCove>",
      "args": ["--mcp", "--config", "<generated config path>",
               "--root", "<authorized dataset folder>"]
    }
  }
}

This is an illustrative shape, not a ready-to-paste configuration. Copy the app-generated values. Keep authorization tokens out of shared prompts and repositories.

2. Inspect the tool list and schemas

The tools and their JSON schemas in your installed version are the source of truth. Use MCP settings to inspect them, run a test call, and review wire logs and call history.

ToolStart here when…
license_statusYou need to verify local Pro status.
dataset_catalogYou need to discover JSONL, Parquet or Arrow shards and partition context.
dataset_scanYou need bounded projected records, supported predicates and provenance across shards.
database_tablesYou need to choose a SQLite or DuckDB table.
database_read_pageYou need a projected, parameterized page from the selected table.
data_materialize_compressedYou explicitly want a new disk-backed copy with an output budget.

3. Run a small, reviewable query

  1. Call license_status and verify Pro is available.
  2. Discover the permitted source and its schema. For a dataset folder, call dataset_catalog with its path and supported partition selection.
  3. Inspect the dataset_scan schema. Select only necessary columns and request a bounded page with a narrow predicate.
  4. Review typed values and provenance. Follow nextCursor until the scope is complete; an empty page with a cursor is not the end.
  5. Verify a representative finding in the desktop inspector. Keep exports and edits as separate explicit actions.
Suggested prompt: “Inspect this source's schema, then analyze one date partition. Read only necessary fields. Show the scope, limits and source provenance. Do not edit, export or scan other folders.”

4. Review the access boundary

The server checks Pro licensing and MCP authorization separately. Directory roots constrain data access. Start with a narrow root; avoid authorizing an entire home directory or drive.

RowCove processes the file locally. The returned tool result is shared with your chosen AI client. Cloud-backed clients may transmit those returned rows to their model provider. Redact sensitive fields and respect your organization's data policy.

Use the server's explicit confirmations for writes and compressed preparation. Inspect the output budget and destination before approving work. Rotate the MCP token when a client should lose access.

Common connection problems

Pro required
Restore your purchase in the desktop app. Release Preview exposes license status without enabling paid data tools.
Invalid or expired MCP authorization
Copy the current configuration from RowCove after a token rotation and reconnect the client.
Path outside authorized roots
Select the correct dataset folder in the authorization settings. Do not broaden access blindly.
Unexpected or incomplete result
Review the call schema, wire log, continuation cursor and operation limits. Distinguish loaded rows from a full-source query.