jamovi MCP for macOS

English | 简体中文

A local stdio MCP server that lets Claude, Cursor, and other MCP clients control jamovi, on macOS and Windows.

This is a port of yjm110517/jamovi-mcp (Windows-only) to macOS. It also fixes several issues found along the way and adds a tool for changing a column's measure type. See Credits.

Open datasets, inspect schemas, edit cells, run statistical analyses, export results, and save .omv files through a local jamovi engine.

Fastest Setup

Copy this into your MCP client config. Claude Code, Claude Desktop, Cursor, and most stdio MCP clients use this command + args shape:

{
  "mcpServers": {
    "jamovi": {
      "command": "uvx",
      "args": [
        "--from",
        "git+https://github.com/Sanjay-Khadka/jamovi-mcp-macos.git",
        "jamovi-mcp"
      ]
    }
  }
}

If you use Claude Code, do not add "type": "stdio" unless your client documentation explicitly requires it, and do not copy a placeholder JAMOVI_HOME value into your config. Start with the minimal config above.

Optional check:

uvx --from git+https://github.com/Sanjay-Khadka/jamovi-mcp-macos.git -- jamovi-mcp --check

After it prints jamovi: ... and MCP transport: stdio, restart your MCP client, then call jamovi_open with an absolute local data file path.

This is the recommended setup for normal users. You do not need to clone this repository, install a local lib/ directory, or hardcode a machine-specific Python path.

Related MCP server: mcp-clipboard

Requirements

• Windows or macOS (see macOS)

• jamovi installed locally

• uvx available to the MCP client

• A Python 3.10+ runtime. uvx provisions this automatically — no manual Python install needed.

uvx is part of uv, a Python tool runner. In this README it is used so your MCP client can download and run jamovi-mcp from GitHub without cloning the repository or hardcoding a local Python path.

Install uv on Windows:

winget install astral-sh.uv

If your MCP client still cannot find uvx after installation, restart the MCP client and, if needed, restart the terminal or Windows session so PATH changes are visible.

jamovi itself is required because this MCP starts a local jamovi engine process. Python does not need to be installed in any specific directory.

jamovi Version Discovery

No configuration required in most cases. The server scans all fixed drives:

1. *\Program Files\jamovi* and *\Program Files (x86)\jamovi* on every drive

2. *\jamovi* at each drive root (e.g. D:\jamovi 2.6)

3. *\*\jamovi* one level below each drive root (e.g. D:\Tools\jamovi)

The newest valid installation is selected automatically.

On macOS the server looks for /Applications/jamovi*.app and ~/Applications/jamovi*.app instead.

When auto-discovery cannot find jamovi

If jamovi is installed deep inside a custom folder tree, set JAMOVI_HOME in the MCP config:

{
  "mcpServers": {
    "jamovi": {
      "command": "uvx",
      "args": [
        "--from",
        "git+https://github.com/Sanjay-Khadka/jamovi-mcp-macos.git",
        "jamovi-mcp"
      ],
      "env": {
        "JAMOVI_HOME": "D:\\MyTools\\jamovi"
      }
    }
  }
}

JAMOVI_HOME must point to the directory containing Frameworks and Resources. To verify you have the right path, run the directory listing and confirm you see both folders:

dir "D:\MyTools\jamovi"
# Should show: Frameworks\  Resources\  bin\  ...

macOS

macOS uses the installed jamovi app; nothing else from jamovi needs configuring.

• Install jamovi from jamovi.org into /Applications (or ~/Applications).

• For the uvx config above, install uv with brew install uv.

• JAMOVI_HOME is optional. If you set it, use the app bundle or its Contents folder; both work:

"env": { "JAMOVI_HOME": "/Applications/jamovi.app/Contents" }

To run from a local clone instead (see Development Install), point the client at the virtualenv's executable. For Claude Desktop, edit ~/Library/Application Support/Claude/claude_desktop_config.json and restart Claude Desktop:

{
  "mcpServers": {
    "jamovi": {
      "command": "/absolute/path/to/jamovi-mcp-macos/.venv/bin/jamovi-mcp",
      "env": { "JAMOVI_HOME": "/Applications/jamovi.app/Contents" }
    }
  }
}

How it runs: the server starts jamovi's bundled Python (Contents/MacOS/python) the same way the jamovi app does, with the environment from Contents/Resources/env.conf and Contents/MacOS as the working directory. This engine is separate from any jamovi window you have open. Quitting the MCP client stops the engine and its R processes.

Tested on macOS 27.0 (Apple Silicon) with jamovi 2.7.39.0, Python 3.13 and mcp 1.30.0. Intel Macs have not been tested.

Example Workflow

You can ask your MCP client to:

Open survey.csv, show the variables, read the first 10 rows, run a t-test, and save the project as analysis.omv.

Typical tool sequence:

1. jamovi_open

2. jamovi_get_schema

3. jamovi_get_data

4. jamovi_run_analysis

5. jamovi_save

MCP Tools

This server exposes 11 MCP tools.

Usage Examples

Open a CSV file:

{
  "file_path": "C:\\Users\\you\\data\\example.csv"
}

Read the active dataset schema:

{}

Read the first 10 rows and first 3 columns:

{
  "row_start": 0,
  "row_count": 10,
  "column_start": 0,
  "column_count": 3
}

Set a single cell value:

{
  "row": 0,
  "column": 1,
  "value": 10
}

Save the active dataset:

{
  "file_path": "C:\\Users\\you\\data\\output.omv",
  "overwrite": true
}

Run an analysis:

{
  "ns": "jmv",
  "name": "ttestIS",
  "options": {
    "vars": ["score"],
    "students": true
  },
  "analysis_id": 2
}

Architecture

flowchart LR
    Client["MCP Client"] --> Stdio["stdio MCP transport"]
    Stdio --> Server["jamovi_mcp.server"]

    Server --> ToolMap["Tool dispatcher"]
    ToolMap --> FileTools["tools.files"]
    ToolMap --> DataTools["tools.data"]
    ToolMap --> AnalysisTools["tools.analysis"]

    FileTools --> Connection["JamoviConnection"]
    DataTools --> Connection
    AnalysisTools --> Connection

    Server --> Engine["EngineManager"]
    Engine --> Config["config.py"]
    Config --> Discovery["JAMOVI_HOME or Program Files discovery"]
    Config --> EnvConf["bin/env.conf parsing"]
    Discovery --> JamoviInstall["Local jamovi installation"]
    EnvConf --> JamoviInstall

    Engine --> JamoviServer["jamovi.server subprocess"]
    JamoviInstall --> JamoviServer

    Connection --> HTTP["HTTP open/save endpoints"]
    Connection --> WS["WebSocket + protobuf coms"]
    HTTP --> JamoviServer
    WS --> JamoviServer

    AnalysisTools --> Registry["analyses.py registry"]
    Registry --> Modules["Resources/modules YAML"]
    Modules --> JamoviInstall

At startup, EngineManager selects a jamovi installation through config.py, builds the process environment from jamovi's own bin/env.conf, and launches jamovi.server. The MCP server connects to that local engine through JamoviConnection. File operations use jamovi's HTTP routes, while dataset and analysis operations use WebSocket messages encoded with the bundled protobuf definitions.

Compatibility

Verified locally:

• Windows: Python 3.10 - 3.12, jamovi 2.6.19.0 (before the macOS changes; the end-to-end workflow has not been re-run on Windows since)

• macOS 27.0 (Apple Silicon): Python 3.13, jamovi 2.7.39.0

Designed compatibility:

• Any Python 3.10+ runtime (the minimum is enforced by the MCP SDK).

• Optional version selection through JAMOVI_HOME.

• Automatic newest-version selection when multiple jamovi* directories are installed under standard Program Files locations.

Known limitations:

• Exports contain tables, notes and text. Plots are listed by title but not embedded.

• jamovi_get_analysis and jamovi_export_results only see analyses run in the current session. After opening a saved .omv, run the analysis again to read or export it.

• Results live in the MCP's own jamovi engine. Open the saved .omv in jamovi to see them in the app.

• The server uses the MCP Python SDK 1.x API, so mcp is pinned below 2.0.

• If a future jamovi release changes jamovi.proto, the WebSocket request types, or the HTTP open/save routes, this MCP may need an adapter update and regenerated protobuf code.

Troubleshooting

uvx is not found

Install uv so your MCP client can run uvx, then restart the MCP client:

winget install astral-sh.uv

uvx means "run a Python tool through uv". If you do not want to use uvx, use the development install below and configure the installed jamovi-mcp command instead.

jamovi-mcp requires Python 3.10 or newer

Your MCP client is using an older Python runtime. With uvx, uv automatically provisions the correct Python version — no manual setup required. If you manage Python yourself, point the MCP command to a Python 3.10+ executable:

{
  "command": "C:\\Path\\To\\Python\\python.exe",
  "args": ["-m", "jamovi_mcp"]
}

This is an advanced fallback. It is not the recommended setup and the path will differ on every computer.

Invalid JAMOVI_HOME

JAMOVI_HOME must point to the jamovi installation directory that contains Frameworks and Resources.

Example:

{
  "env": {
    "JAMOVI_HOME": "C:\\Your\\jamovi\\Install\\Path"
  }
}

Do not copy C:\\Your\\jamovi\\Install\\Path literally. It is a placeholder. Replace it with a real path, or remove the entire env block when jamovi is installed under a standard Program Files location.

jamovi is installed but not detected

Set JAMOVI_HOME explicitly in the MCP client config. This is also recommended when testing a specific jamovi version.

Claude Code reports an MCP config schema error

Start with the minimal config and keep only command and args:

{
  "mcpServers": {
    "jamovi": {
      "command": "uvx",
      "args": [
        "--from",
        "git+https://github.com/Sanjay-Khadka/jamovi-mcp-macos.git",
        "jamovi-mcp"
      ]
    }
  }
}

The most common causes are extra fields unsupported by the current client, or copying a placeholder JAMOVI_HOME value into the config. Get the minimal config working first, then add a real env block only if needed.

File open or save fails

Use absolute paths (~ is expanded; relative paths are resolved against the MCP server's working directory) and make sure the user running the MCP client has permission to read or write that location. For save operations, pass "overwrite": true if the target file already exists.

Analysis tools return unexpected results

First call jamovi_list_analyses, then jamovi_get_analysis_options for the target analysis. jamovi analysis option schemas are module-specific and can differ between versions or installed modules.

Development Install

Normal users should use the uvx MCP config above. Clone the repository only if you want to develop or test the code locally.

git clone https://github.com/Sanjay-Khadka/jamovi-mcp-macos.git
cd jamovi-mcp-macos
py -3.10 -m pip install -e .

If your system does not have the Windows Python launcher, use any Python 3.10+ executable instead:

python -m pip install -e .

Run tests:

py -3.10 -m pytest -q

On macOS:

python3 -m venv .venv
.venv/bin/pip install -e . pytest pytest-asyncio
.venv/bin/python -m pytest -q
.venv/bin/jamovi-mcp --check

Start the MCP server directly:

py -3.10 -m jamovi_mcp

Important source areas:

• src/jamovi_mcp/server.py: MCP server and tool registration.

• src/jamovi_mcp/engine.py: jamovi engine subprocess lifecycle.

• src/jamovi_mcp/config.py: jamovi install discovery and environment setup.

• src/jamovi_mcp/layout.py: where jamovi keeps its files on Windows and macOS.

• src/jamovi_mcp/results_export.py: renders analysis results as text or HTML.

• src/jamovi_mcp/connection.py: HTTP, WebSocket, and protobuf communication.

• src/jamovi_mcp/tools/: MCP tool implementations.

• src/jamovi_mcp/analyses.py: analysis registry built from jamovi module YAML files.

• tests/: unit tests for data conversion, save handling, config discovery on Windows and macOS, engine process handling, and results export.

Do not commit lib/ or other local dependency target directories. Install dependencies through pyproject.toml.

Security Notes

This MCP starts a local jamovi process and reads or writes local files whose paths are provided through MCP tool calls.

• The engine is started locally and connected through 127.0.0.1.

• File paths are supplied by the MCP client/user.

• Do not expose this server to untrusted clients.

• Do not pass sensitive data files to an MCP client you do not trust.

• Do not commit private local config, access tokens, API keys, or datasets.

Roadmap

• Add GitHub Actions CI.

• Add broader integration tests across more jamovi versions.

• Improve structured parsing for analysis result payloads.

• Add more explicit typed response schemas for each MCP tool.

• Document common jamovi analysis recipes.

Contributing

Pull requests are welcome. Please keep changes focused, run the test suite before submitting, and include tests for behavior changes.

For compatibility work, include the jamovi version, operating system version, and Python version used for testing.

Repository Contents

Files that should be committed:

• README.md

• README.zh-CN.md

• LICENSE

• .gitignore

• pyproject.toml

• docs/

• src/

• tests/

Files and directories that should not be committed:

• lib/

• .pytest_cache/

• .ruff_cache/

• __pycache__/

• local CSV/OMV/log/tmp files

• private local config, tokens, and API keys

Credits

Based on jamovi-mcp by yjm110517, used under the MIT License. The original project's commit history is kept in this repository.

Changes in this port:

• macOS support: discovery of jamovi.app, the bundle's layout and env.conf, and launching the bundled Python the way the jamovi app does.

• Engine process handling: its own stdin pipe (it no longer shares the MCP stdio transport), continuous output reading, startup timeout, and clean shutdown with no orphaned R processes.

• jamovi 2.7 support: analysis option schemas from jamovi-full.yaml, and protobuf code regenerated from jamovi 2.7.39.

• jamovi_get_analysis and jamovi_export_results now return real results; export renders text or HTML.

• New jamovi_set_measure_type tool, ~ path support, and more tests.

License

MIT. See LICENSE.