RxInfer FastMCP Example

An MCP server that exposes Bayesian house price estimation powered by RxInfer.jl. Point it at a CSV of housing data, and it learns a probabilistic pricing model that predicts prices with honest uncertainty -- when information is missing, the model says "I'm less sure" instead of silently guessing.

Built with FastMCP, JuliaCall, and uv.

Tools

Tool	Description
run_bayesian_analysis	Load a CSV file and train the Bayesian regression model
describe_data	Inspect the loaded dataset -- columns, stats, and what the model recognizes
estimate_house_price	Predict price for a house (any features can be omitted)

The CSV should have a price column and any of: bedrooms, bathrooms, living_area, floors, basement_area, yr_built, label.

Installation

uv sync

Julia itself and all Julia dependencies are managed automatically by juliapkg: they are pinned in juliapkg.json and installed into an isolated Julia environment under .venv/julia_env (independent of any global Julia setup). The first server start downloads and precompiles the full stack (RxInfer, Plots, …), which can take several minutes; subsequent starts use the compiled cache. If you hit issues, see the juliapkg docs.

Usage

Start the server in the background before connecting any MCP client:

uv run fastmcp run main.py:mcp --transport sse --port 8001

The first startup is slow (~30s) because Julia compiles packages and generates a test CSV. Subsequent runs use the compiled cache and start much faster. Wait until you see the server listening message before proceeding.

Test dataset

Starting the server writes a synthetic dataset to test_houses.csv in the project root (~20,000 houses). This is the dataset used in the demo/presentation — there's no need to keep it around, since it's regenerated on demand.

• Deterministic: it's generated with a fixed seed (StableRNG(123)), so every regeneration produces the exact same dataset. Reproducible demos.
• Gitignored on purpose: to get it back, just (re)start the server. Delete the file first if you want a guaranteed-fresh copy.
• Columns: bedrooms, bathrooms, living_area, floors, basement_area, yr_built, label, price.

To use it in the demo, point run_bayesian_analysis at test_houses.csv, or drag-and-drop the file into Claude Desktop.

The generation logic lives in rxinfer.jl and reuses the random_house / price_house functions from example.jl. To skip it on startup (faster boot when you already have a dataset), set SKIP_TEST_CSV=true:

SKIP_TEST_CSV=true uv run fastmcp run main.py:mcp --transport sse --port 8001

Pluto Notebook

The Bayesian model is developed in example.jl, a Pluto.jl reactive notebook. To explore or modify the model interactively, install Pluto following the instructions on the website, then open the notebook:

using Pluto; Pluto.run()

This opens a browser-based environment where you can open example.jl, tweak the model, and see results update in real time.

Claude Desktop

To connect via SSE in Claude Desktop, add the following to your Claude Desktop MCP configuration:

{
  "mcpServers": {
    "rxinfer": {
      "command": "npx",
      "args": [
        "mcp-remote@latest",
        "http://localhost:8001/sse"
      ]
    }
  }
}

Make sure the server is running before launching Claude Desktop.

Claude Code

A .mcp.json is included that connects to the server on localhost:8001. Start the server first in a separate terminal:

uv run fastmcp run main.py:mcp --transport sse --port 8001

Then launch Claude Code in the project directory. The server will be detected automatically — approve it when prompted.

Example conversation

Ask Claude to run Bayesian analysis on the test dataset (test_houses.csv, created automatically when starting the server — see Test dataset)

Prompt it to estimate a price for some house, some features might be missing

Think of the situation where you actually got the missing data about the previously unknown features and check the new result