Skip to content
ReviewRamble

LocalData MCP

Connects an agent to local files and databases and gives it real statistics tools: tests, regression, clustering and forecasting.

Reviewed by Sheela · Updated

Setup docs Open source

Our rating

3.5/5

Lets an agent run the statistics instead of estimating them, from a capable project almost nobody uses yet.

What it is

LocalData MCP connects AI agents to databases and data files and adds a data-science toolkit. It reads CSV, TSV, JSON, XML, INI, Parquet, Feather, Arrow and spreadsheets, and connects to SQLite, PostgreSQL, MySQL and DuckDB, with optional extras for SQL Server, MongoDB, CouchDB, key-value stores, InfluxDB and Neo4j. Queries stream in chunks within a 2 GB memory limit by default.

Its 12 data-science tools cover hypothesis tests, regression, clustering, anomaly detection, dimensionality reduction and ARIMA or ETS time-series forecasting, each taking a connection name and a SQL query. The repo also works as a Claude Code plugin with 18 skills and 11 agents. Data-science dependencies download about 200 MB on first use. Apache 2.0 with 12 stars; v2.0.0 shipped on 6 April 2026.

Install

uvx localdata-mcp

Our review

Ask a chatbot whether two regions' sales really differ and it will reason its way to a plausible answer. LocalData MCP lets the agent run the test instead: analyze_hypothesis_test takes a SQL query and a grouping column and returns an actual result.

The design is consistent: every analysis tool takes a connection name, a query and column names, so there is no separate loading step. Coverage is wide, from regression with R² and RMSE to clustering, anomaly detection and ARIMA or ETS forecasts, and the bundled Claude Code skills and agents (an ml-analyst, a forecaster) show the intended workflows.

Safety defaults are sensible. File access is confined to the working directory unless you configure allowed paths, and queries stream within a memory budget. The README does not say whether execute_query is limited to read-only SQL, so connect with a read-only database user.

The catch is maturity. It has 12 stars, 52 open issues and a single maintainer, and the first run downloads about 200 MB of scientific libraries. The statistics come from scipy, scikit-learn and statsmodels, which is reassuring; choosing the right test is still up to you.

Where it earns its place

  • An operations analyst at a 40-store retailer

    Check whether a promotion changed sales and forecast next quarter

    Connect the sales CSV, run a hypothesis test by store group, then forecast_time_series on weekly totals. Free; runs locally.

  • A backend developer investigating payment anomalies

    Flag unusual transactions in a Postgres table

    Connect with a read-only user and run anomaly detection over a query of recent transactions.

Before you commit

Setup
uv tool install localdata-mcp, then add localdata-mcp (or uvx localdata-mcp) to the MCP config. First use downloads about 200 MB of data-science libraries.
Integrations
SQLite, PostgreSQL, MySQL, DuckDB and common file formats; optional extras for SQL Server, MongoDB, InfluxDB, Neo4j and more; Claude Code plugin.
Data and privacy
Runs locally; file access restricted to allowed paths, the working directory by default. Apache 2.0.
Limits
2 GB default memory budget; 12 stars and 52 open issues; SQL write restrictions not documented.

What's good

  • Real statistical tests, regression, clustering and forecasting
  • Wide range of files and databases
  • File access confined to the working directory by default
  • Claude Code plugin with analysis skills and agents

What's not

  • Tiny user base, 52 open issues
  • Read-only SQL enforcement not documented
  • About 200 MB of dependencies on first use

Best for

Analysts and developers who want an agent to run real statistics on local data.

Alternatives

Filesystem (reference server)

Model Context ProtocolDeveloper

4.1/5

Official MCP reference server for reading, writing, searching and moving files inside folders you allow.

LocalData MCP: common questions

Is LocalData MCP worth it?
Lets an agent run the statistics instead of estimating them, from a capable project almost nobody uses yet. We rate it 3.5 out of 5.
How much does LocalData MCP cost?
Free and open source under Apache 2.0.
What are the downsides of LocalData MCP?
Tiny user base, 52 open issues; read-only SQL enforcement not documented; about 200 MB of dependencies on first use.
Who is LocalData MCP best for?
Analysts and developers who want an agent to run real statistics on local data.
Is LocalData MCP safe for company data?
Runs locally; file access restricted to allowed paths, the working directory by default. Apache 2.0.
What are the best LocalData MCP alternatives?
The closest alternatives we have reviewed are Filesystem (reference server).

Answers checked Oct 24, 2026.