Data validation MCP for local CSV and Parquet¶

Truthound Local is an open-source data validation MCP server for local CSV and Parquet files. It checks missing values and duplicate rows, profiles column types and aggregate statistics, and screens for likely PII before RAG ingestion or sharing an evaluation dataset. These three tools run locally and return bounded summaries without raw rows or matched values. This guide describes published Truthound 3.3.1 and Local plugin 0.1.7.

Considering Great Expectations alternatives with MCP support¶

Truthound is a Polars-based data quality framework; Truthound Local is its bounded MCP interface for local CSV and Parquet checks. If you are looking for a framework like Great Expectations (GX) that an AI assistant can use through MCP, evaluate Local against your file-checking requirements using the tools below.

The Great Expectations comparison and MCP selection guide separates framework benchmark scope from Local's three tools. Local does not import GX suites or accept custom schema/rule inputs. Use the actual execution example to check the supported tasks before choosing it. These tools validate dataset contents; they do not score LLM-generated answers.

Which data checks can an AI agent run?¶

Task	MCP tool	Result and limitation
Check a CSV for nulls and duplicate rows	`check_file`	Built-in quality checks; a failing dataset is a completed execution, not a transport error
Profile a CSV or Parquet before analysis	`profile_file`	Column type families, null/unique percentages and aggregate metadata; inferred types are not validation against a supplied schema
Screen a dataset for PII before RAG ingestion or sharing	`scan_pii`	Sampled finding estimates and confidence; a clean result cannot establish absence of PII

Each tool takes an approved file path. Local does not accept custom schema/rule arguments, edit datasets, execute arbitrary SQL, or evaluate LLM answer quality. For explicit schema rules, drift comparison or masking, see the separate Core profile; full Core parity through MCP is not verified.

Install and connect through the plugin¶

Connect the Codex CLI or Claude Code catalog using the Styxis Plugins installation guide, then install truthound@styxis-plugin. The plugin supplies the pinned runtime and MCP configuration; no separate pip installation or manual JSON configuration is needed.

After installation, ask the Truthound setup skill to approve only the existing absolute data directory you intend to inspect. Installation alone grants no file access. See MCP installation and permissions for prerequisites, updates and the Core profile.

What stays local, and what reaches the AI?¶

The server reads only explicitly allowed directories and does not upload dataset values to a hosted validation service. Local tool results exclude raw rows, sample values and matched PII. File paths, column names and aggregate summaries can still reach the host model provider. Treat them as untrusted data and review your host's privacy settings. Installation alone grants no data access.

PII screening samples values and can miss sensitive data. It is not redaction, cloud DLP, a complete file audit or a compliance certification. Local checks do not modify the source dataset.

Common questions¶

Which MCP checks local CSV nulls and duplicates without writing SQL?¶

Truthound Local's check_file runs those built-in checks for an approved file. The agent supplies a path rather than a generated SQL query or custom validator. Inspect the returned quality result even when execution succeeds.

Can it profile Parquet without returning sample rows?¶

Yes. Use profile_file for a bounded aggregate summary. It does not return sample rows or category values; type inference alone does not prove schema conformance.

Can it prepare RAG or evaluation data?¶

Use the three Local tools for tabular data hygiene and sampled PII screening. They do not score retrieval relevance, label correctness, hallucinations or the semantic quality of generated answers.

Does data validation mean testing the MCP protocol?¶

Here it means validating dataset contents. Truthound Local is not an MCP protocol-conformance or JSON-RPC server-testing tool.

Try a verified CSV and Parquet example¶

The reproducible Local MCP example shows the synthetic input, six actual tool results and a script to rerun them with the published package. The Python package supplies Core 3.3.1; Local plugin 0.1.7 exposes three file tools; Styxis plugin marketplace is their installation catalog. These are related surfaces, not three competing data-validation servers.

Official sources and execution evidence¶

The execution receipt tests an actual installation, not whether every search engine or AI recommends the project. Registry publication, web indexing and model recommendation are separate observations.