MCP R Playground

An MCP server that allows AI models to execute R code, see its results, and draw and observe plots. It can be used for sophisticated agentic deployments, but also as a way to augment AI clients like Claude Desktop when talking to them about scientific papers.

Features:

• Stateful sessions: each conversation thread gets a new session, but the session can persist across calls and user/assistant interactions.
• Graphics output: multimodal models can draw plots using standard R libraries like ggplot, see those plots, and react to them.
• Works in all common operating systems/architectures - Windows x64 / arm64, MacOS, Linux.
• When running locally, __NO HOST ISOLATION__: while each session runs as a separate R environment, they have access to global dependencies and all files on the computer. While unlikely, a rogue model could write R code that deletes your important files.
  • If you need host isolation, you should run this MCP in Docker. Instructions for that are provided below.

Configuration

Currently there’s just one configuration parameter that can be set as an environment variable:

• RPLAYGROUND_MCP_SUPPORT_IMAGE_OUTPUT, default True. If set to False, image output will be disabled, and tool descriptions will be made to reflect that.

Usage

By default, the AIs will have access to all globally installed R packages, and can install whatever package they want. These installations will persist. You can pre-install important packages to make them available in advance.

Installation

Basic instructions for technical users:

1. Have R installed, and the R_HOME environment variable set
2. Have a recent version of the uv installed
3. run uvx --python=3.13 rplayground-mcp, and it should just work.

Detailed Installation

This section is for less technical users who want to set up this MCP to use with Claude Desktop or similar AI user interfaces that support MCP extensions.

Windows

• Make sure you’ve set up Claude Desktop to create the MCP configuration file by following instructions here: https://modelcontextprotocol.io/quickstart/user
• Make sure you have R installed. The recommended source is here https://cran.rstudio.com/ .
• Make sure you have uv installed. uv is the project management tool for Python, the programming language this tool is written in. More detailed instructions can be found here https://docs.astral.sh/uv/getting-started/installation/#pypi, we provide the instructions for the most straightforward method:
  1. Open the Terminal app
  2. In the terminal, paste in the following installation command: powershell -ExecutionPolicy ByPass -c "irm https://astral.sh/uv/install.ps1 | iex"
  3. Close the Terminal app and reopen it
  4. type in uv and confirm you don’t see any red errors.
• We have provided a helper script that you can use to set up the MCP server to work with Claude Desktop. You can run it with uv run --python=3.13 https://raw.githubusercontent.com/zygi/r-playground-mcp/refs/heads/master/scripts/setup_helper.py. With your permission, it will:
  • Set the R_HOME environment variable to your R installation
  • Install the MCP inside your Claude Desktop configuration.
• That’s it! Starting Claude Desktop should now display the tools as available. Or just ask it to “test out the execute_r_command tool”.

MacOS

• Make sure you’ve set up Claude Desktop to create the MCP configuration file by following instructions here: https://modelcontextprotocol.io/quickstart/user
• Make sure you have R installed. The recommended source is here https://cran.rstudio.com/ .
• Make sure you have uv installed. uv is the project management tool for Python, the programming language this tool is written in. More detailed instructions can be found here https://docs.astral.sh/uv/getting-started/installation/#pypi, we provide the instructions for the most straightforward method:
  1. Open the Terminal app
  2. In the terminal, paste in the following installation command: curl -LsSf https://astral.sh/uv/install.sh | sh
  3. Close the Terminal app and reopen it
  4. type in uv and confirm you don’t see any red errors.
• We have provided a helper script that you can use to set up the MCP server to work with Claude Desktop. You can run it with uv run --python=3.13 https://raw.githubusercontent.com/zygi/r-playground-mcp/refs/heads/master/scripts/setup_helper.py. With your permission, it will:
  • Set the R_HOME environment variable to your R installation
  • Install the MCP inside your Claude Desktop configuration.
• That’s it! Starting Claude Desktop should now display the tools as available. Or just ask it to “test out the execute_r_command tool”.

Installation (Docker)

We also provide Dockerfiles to run this MCP in an isolated context. This only supports the platform linux/amd64. The image uses the r2u project to make precompiled CRAN packages available. The Dockerfile comes with two versions:

• Slim: docker build --platform=linux/amd64 .. This sets up the r2u repositories but installing a new package from the R session will still take ~10s.
• Fat: docker build --build-arg PREINSTALL_PACKAGES=true --platform=linux/amd64 .. This preinstalls a big selection of packages to the Docker image, and loading them in an MCP session becomes instant.

You are also welcome to edit the Dockerfile and preinstall just the packages relevant to your workflow.

Issues and Contributions

Feel free to create an Issue if you have questions or requests. Small PRs are welcome anytime, larger PRs should be discussed by creating an Issue before a PR is started.