archetana/pysr-mcp-server

PySR MCP Server - Quick Start

What is PySR?

PySR discovers mathematical equations from data using evolutionary algorithms. Instead of black-box models, it finds interpretable formulas like y = 2*x + 1 or y = sin(x) * exp(-x).

Use cases: Scientific discovery, physics modeling, feature engineering, pattern discovery.

Installation

# Install dependencies
pip install fastmcp pysr numpy pandas matplotlib seaborn sympy psutil

# Install Julia backend (one-time, takes 5-10 minutes)
python -c "import pysr; pysr.install()"

Setup

1. File Structure

your-project/
├── pysr_mcp_server.py
└── proxy_server.py

2. Configure Claude Desktop

macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
Windows: %APPDATA%\Claude\claude_desktop_config.json

{
  "mcpServers": {
    "pysr": {
      "command": "python",
      "args": ["/absolute/path/to/proxy_server.py"]
    }
  }
}

⚠️ Use absolute path! Example: /Users/yourname/projects/pysr/proxy_server.py

Running

Open 2 terminals:

Terminal 1 (Main Server):

python pysr_mcp_server.py

Terminal 2 (Proxy):

python proxy_server.py

Then restart Claude Desktop.

Test Prompts

Test 1: train_model_complete (All-in-One)

Test the PySR server with train_model_complete tool.

CSV data:
x,y
1,2
2,4
3,6
4,8
5,10

Parameters:
- target_column: "y"
- job_id: "test_001"
- Use default settings

Show me the discovered equation and metrics.

Test 2: create_model + fit_model (Step-by-Step)

Step 1: Create a PySR model using create_model tool:
- model_id: "custom_001"
- binary_operators: ["+", "-", "*", "/"]
- unary_operators: ["sin", "cos"]
- niterations: 40
- populations: 15

Step 2: Fit this model using fit_model tool with data:
X = [[1], [2], [3], [4], [5]]
y = [1, 4, 9, 16, 25]
variable_names: ["x"]

Show me the equations discovered.

Test 3: Predict with Trained Model

After training model "test_001" from Test 1:

1. Use predict tool to predict for new data:
   X = [[6], [7], [8], [9], [10]]

2. Show predictions and compare with expected pattern (y = 2*x)

Test 4: Multiple Features

Use train_model_complete for multi-variable data:

CSV:
x1,x2,y
1,1,3
2,1,5
1,2,4
2,2,6
3,1,7
3,2,8

Parameters:
- target_column: "y"
- job_id: "multi_001"
- niterations: 50

Discover the equation relating x1, x2 to y.

Test 5: Polynomial Relationship

Train a model to discover this cubic pattern:

CSV:
x,y
1,1
2,8
3,27
4,64
5,125

Parameters:
- target_column: "y"
- job_id: "cubic_001"
- binary_operators: ["+", "-", "*", "/", "^"]
- niterations: 60

What equation does it find?

Test 6: Model Management

Test model lifecycle:

1. list_models - Show all active models
2. get_equations for "model_test_001"
3. export_equation in LaTeX format
4. save_model to disk as "my_model.pkl"
5. delete_model from memory
6. load_model from "my_model.pkl" with new id "reloaded_001"

Test 7: Exponential Pattern

Find equation for exponential growth:

CSV:
t,value
0,1
1,2.718
2,7.389
3,20.086
4,54.598

Parameters:
- target_column: "value"
- unary_operators: ["exp", "log", "sin", "cos"]
- job_id: "exp_001"
- niterations: 50

Hint: This is e^t pattern.

Test 8: Trigonometric Pattern

Discover sine wave equation:

CSV:
x,y
0,0
1.57,1
3.14,0
4.71,-1
6.28,0

Parameters:
- target_column: "y"
- unary_operators: ["sin", "cos", "tan"]
- job_id: "trig_001"

What equation does it find?

Available Tools

Common Parameters

train_model_complete key parameters:

• job_id (required): Unique identifier
• data (required): CSV string or file path
• target_column (required): Column name to predict
• feature_columns: List of input columns (optional)
• niterations: Training iterations (default: 40)
• populations: Number of populations (default: 15)
• binary_operators: ["+", "-", "*", "/", "^"]
• unary_operators: ["sin", "cos", "exp", "log", "sqrt"]
• maxsize: Max equation complexity (default: 20)
• parsimony: Simplicity preference (default: 0.0032)

Troubleshooting

Server won't start (port busy):

lsof -ti:8000 | xargs kill -9

Claude can't connect:

• Check proxy running: ps aux | grep proxy_server
• Use absolute path in config
• Restart Claude Desktop

Training too slow:

• Reduce niterations to 20-30
• Reduce populations to 8-10

Import error:

pip install --upgrade fastmcp pysr numpy pandas sympy psutil

Resources

• PySR GitHub: https://github.com/MilesCranmer/PySR/tree/master/pysr
• PySR Docs: https://astroautomata.com/PySR/
• PySR Paper: https://arxiv.org/abs/2305.01582
• MCP Spec: https://modelcontextprotocol.io/

Quick Reference

# Start servers
python pysr_mcp_server.py  # Terminal 1
python proxy_server.py      # Terminal 2

# Check health
curl http://localhost:8000/health

# Common operators
# Basic: +, -, *, /
# Powers: ^, sqrt
# Trig: sin, cos, tan
# Exp: exp, log

Happy Equation Discovery! 🔬