R MCP Server

An MCP (Model Context Protocol) server that lets AI assistants execute R code, create visualizations, analyze data, and manage packages — all through a local Rscript CLI.

Features — 62 Tools

Execution (3 tools)

Visualization (5 tools)

Statistical Analysis (5 tools)

Data Wrangling (5 tools)

Time Series (4 tools)

Clustering (2 tools)

Advanced Statistics (7 tools)

Interactive & Publication Plots (5 tools)

Probability Distributions (5 tools)

Proportion & Contingency Tests (5 tools)

Regression & Post-hoc (6 tools)

Exploratory Data Analysis (5 tools)

Utilities (5 tools)

Prerequisites

• R (>= 4.0) with Rscript on your PATH
• Python (>= 3.10)

Install R from CRAN or via Homebrew:

brew install r

Installation

git clone https://github.com/sergiudanstan/r-mcp.git
cd r-mcp
pip install -e .

Usage

With Claude Code

Add to your Claude Code MCP settings (~/.claude/settings.json):

{
  "mcpServers": {
    "r": {
      "command": "python",
      "args": ["-m", "r_mcp"],
      "cwd": "/path/to/r-mcp"
    }
  }
}

Standalone

python -m r_mcp

The server communicates over stdio using the MCP protocol.

How It Works

The server wraps the Rscript --vanilla CLI. Each tool call spawns a fresh R session, executes the code, and returns structured JSON results. Code is wrapped in tryCatch for clean error reporting.

• Workspace: Output files (plots, rendered docs) are saved to ~/r-mcp-workspace/
• Timeout: Default 60s per execution (configurable per call)
• Safety: Path traversal prevention on file outputs; output truncation at 50K chars

Examples

Run R code

# Via the evaluate_r_code tool
x <- rnorm(100)
cat("Mean:", mean(x), "\nSD:", sd(x), "\n")

Create a plot

# Via the create_r_plot tool
library(ggplot2)
df <- data.frame(x = rnorm(200), y = rnorm(200))
ggplot(df, aes(x, y)) + geom_point(alpha = 0.5) + theme_minimal()

Probability distributions

# Via the distribution_calculator tool
# Compute P(X <= 1.96) for standard normal
pnorm(1.96, mean=0, sd=1)

# Via the distribution_plot tool
# Visualize chi-squared(5) distribution with 1000 samples

Hypothesis testing

# Via the proportion_test tool
# Test if 42 out of 100 differs from 50%
prop.test(42, 100, p = 0.5)

# Via the hypothesis_test tool
# Two-sample t-test
t.test(x, y, alternative = "two.sided")

Analyze a CSV

Use get_data_summary with a file path to get dimensions, column types, summary statistics, and a preview.

License

MIT