TabDat Model Context Protocol (MCP) Server¶

TabDat includes a built-in Model Context Protocol (MCP) server conforming to the JSON-RPC 2.0 stdio protocol (protocolVersion: "2024-11-05"). This enables AI coding agents and LLM applications (such as Claude Desktop, Cursor, Google Antigravity, Goose, Cline, and Windsurf) to directly query, explore, transform, and model tabular datasets with native Stata-inspired command semantics.

1. Quickstart¶

Launch via CLI¶

Run the MCP server over standard I/O streams:

tabdat --mcp

Or using the dedicated entry point:

tabdat-mcp

2. Configuration for AI Clients¶

Claude Desktop¶

Add TabDat to your claude_desktop_config.json (located at ~/Library/Application Support/Claude/claude_desktop_config.json on macOS or %APPDATA%\Claude\claude_desktop_config.json on Windows):

{
  "mcpServers": {
    "tabdat": {
      "command": "uvx",
      "args": ["tabdat-explore", "--mcp"]
    }
  }
}

If developing locally:

{
  "mcpServers": {
    "tabdat": {
      "command": "uv",
      "args": ["run", "--directory", "/path/to/tabdat-explore", "tabdat", "--mcp"]
    }
  }
}

Cursor¶

Add TabDat to your Cursor MCP settings (.cursor/mcp.json or Global Cursor Settings -> MCP):

{
  "mcpServers": {
    "tabdat": {
      "command": "tabdat",
      "args": ["--mcp"]
    }
  }
}

Google Antigravity / Gemini CLI¶

Add TabDat as an MCP tool provider:

agy mcp add tabdat tabdat --mcp

3. Available Tools¶

Tool Name	Description	Key Arguments
`tabdat_execute`	Execute a single TabDat command in the active dataset session.	`command` (str, required), `output_format` ("terminal" \| "json")
`tabdat_batch`	Execute an ordered list of TabDat commands sequentially.	`commands` (list[str], required), `output_format` ("terminal" \| "json")
`tabdat_script`	Run an entire TabDat script file or inline `.td` script content.	`script_content` (str), `file_path` (str), `output_format`
`tabdat_status`	Inspect live session state (table name, row count, backend engine, model).	`output_format` ("terminal" \| "json")
`tabdat_describe_command`	Look up schema, arguments, options, and syntax for any command.	`command_name` (str, required)
`tabdat_list_commands`	List all supported commands and their declared side-effects.	None
`tabdat_get_help`	Fetch packaged markdown documentation for any command or topic.	`topic` (str, required)
`tabdat_explain`	Parse and validate syntax of a command without running it.	`command` (str, required)
`tabdat_doctor`	Check backend availability (DuckDB, Arrow, Polars, Statsmodels, etc.).	None
`tabdat_reset_session`	Clear the active session and reset all loaded tables and models.	None

Tool Execution Example¶

An agent calling tabdat_execute:

{
  "name": "tabdat_execute",
  "arguments": {
    "command": "use data/synthetic.parquet",
    "output_format": "terminal"
  }
}

Followed by modeling:

{
  "name": "tabdat_execute",
  "arguments": {
    "command": "regress income age bmi, robust",
    "output_format": "terminal"
  }
}

Session state (loaded datasets, computed variables, active tables, last estimation models) is persisted across consecutive tool calls.

4. Live Resources¶

TabDat exposes live dynamic URI resources for real-time state inspection:

Resource URI	Description	MIME Type
`tabdat://session/status`	Current session state, active dataset path, row count, and backend mode.	`application/json`
`tabdat://session/schema`	Column names and structural datatypes of the active dataset.	`application/json`
`tabdat://catalog/commands`	Comprehensive catalog of supported commands and effect categories.	`application/json`

5. Guided Prompt Templates¶

Prompt Name	Description	Arguments
`eda_workflow`	Step-by-step exploratory data analysis template.	`file_path` (required), `focus_variables` (optional)
`econometric_analysis`	Regression modeling and post-estimation diagnostic workflow.	`file_path` (req), `dependent_var` (req), `independent_vars` (req), `estimator`
`data_cleaning`	Variable creation, filtering, recoding, and export template.	`file_path` (req), `task_description` (req)