MCP server

With more than one adapter, dataveil can also expose itself directly as an MCP server, usable outside SQLMesh entirely:

pip install dataveil[duckdb,mcp]   # or [postgres,mcp] / [sqlmesh,mcp]

{
  "mcpServers": {
    "dataveil": {
      "command": "dataveil-mcp",
      "env": { "DATAVEIL_ADAPTER": "duckdb", "DATAVEIL_DUCKDB_PATH": "/path/to/your.duckdb" }
    }
  }
}

DATAVEIL_ADAPTER selects duckdb / postgres / sqlmesh, each with its own setting: DATAVEIL_DUCKDB_PATH (a file path, or :memory:), DATAVEIL_POSTGRES_URL (a SQLAlchemy URL), or SQLMESH_PROJECT_PATH (same env var sqlmesh-mcp uses). Four tools: list_tables, profile, propose_cleansing_plan, apply_cleansing_plan (requires confirm=true, destructiveHint) -- the same three-tool shape as sqlmesh-mcp's dataveil-backed tools, just against table instead of model_name, since there's no SQLMesh Context or physical-snapshot concept here.

The adapter can also come from the [adapter] section of a configuration file, and the server applies the file's [profile] and [classify] thresholds to every call. Environment variables take precedence over the file. Set DATAVEIL_ADAPTER to the name of an adapter plugin to use one.

Tool	Changes data	Does
list_tables	no	Lists the tables or models the adapter can see.
profile	no	Returns the aggregate-only profile of a table.
propose_cleansing_plan	no	Returns the profile, the classification of each column and the operation vocabulary: everything an agent needs to write a plan.
apply_cleansing_plan	yes	Validates and runs a plan. Requires confirm=true.

The server never proposes a plan itself and never calls an LLM. Every call is recorded in the audit log.