Skip to main content

Agents and MCP

Give agents controlled dataset tools or add AI assistance to documentation.

Connect undatum to an MCP client​

pip install "undatum[mcp]"
undatum mcp tools
undatum mcp serve

Add this to Cursor mcp.json or Claude Desktop MCP settings:

{
  "mcpServers": {
    "undatum": {
      "command": "undatum",
      "args": ["mcp", "serve"]
    }
  }
}

Every data operation is a tool (dedup, join, rename, where, ...): without output_path it returns records inline, and writing a file requires confirm=true. Paths stay inside the server's --root. undatum mcp tools --parity shows which CLI commands have a tool. Full catalog and flags: MCP and mcp.

Let the client browse datasets​

The server also lists the data files under its root as resources (undatum://datasets, each file's schema and sample) and offers ready prompts: profile-dataset, draft-validation-rules, plan-conversion and document-dataset. See resources and prompts.

Same results for scripts and agents​

Informational commands print the same versioned JSON documents that the agent tools return:

undatum count data.csv --json
undatum sniff data.csv --json
undatum headers data.csv --json

See JSON output.

Generate assisted dataset documentation​

undatum ai doc data.csv --format-out json --blocks general,schema,quality

Python tools without MCP​

from undatum import tools
from undatum.tools import schemas

print(tools.detect_format("data.csv"))
print(schemas.call_tool("query_sql", {"path": "data.csv", "query": "SELECT * FROM data LIMIT 5"}))
print(schemas.to_openai_functions())

See MCP, AI documentation, ai, and the Python SDK.