python-session-mcp

Run Python in a session that stays alive, and expose it to LLM clients over the Model Context Protocol.

Author: Dr Merwan Roudane

What makes it different

The session persists. A DataFrame loaded in one call is still there in the next, so an analysis is built up in steps rather than resent whole each time.

Your interpreter, not the server's. The server may well be installed under a bare Python with no pandas in it. The interpreter that runs your code is chosen separately: a conda or Anaconda installation is preferred when one is present, and PYTHON_MCP_INTERPRETER overrides that.

A crash costs one process. Code runs in a worker, not in the server. Exhaust memory, call sys.exit, crash a C extension — the worker is replaced and the server carries on. It also keeps user code away from the server's stdin, which under MCP is the JSON-RPC stream itself.

Install

pip install python-session-mcp

Library use

from python_mcp import PythonRunner

with PythonRunner() as py:
    py.run("import pandas as pd, statsmodels.api as sm")
    py.run("df = pd.read_csv('macro.csv')")
    print(py.run("df.describe()"))

    py.run("m = sm.OLS(df['y'], sm.add_constant(df[['x','z']])).fit()")
    print(py.run("m.summary()"))
    print(py.value("m.params.to_dict()"))    # a real Python dict

A final expression is shown the way a REPL would, so df.head() on its own displays the frame without print().

MCP server use

{
  "mcpServers": {
    "python": {
      "command": "python-session-mcp",
      "env": { "PYTHON_MCP_INTERPRETER": "C:\\Users\\you\\anaconda3\\python.exe" }
    }
  }
}

Tools

Session

Tool	Purpose
python_status	Which interpreter, and which packages it actually has
reset_namespace	Forget everything, optionally restarting the interpreter

Running code

Tool	Purpose
run_python	Main tool. Run code in the persistent session
list_names	What is currently defined
describe_object	Type, shape, dtypes and a peek at one object
get_value	Bring a JSON-representable value back

Data

Tool	Purpose
load_data	Read .csv, .xlsx, .dta, .parquet, .sav or .json
save_data	Write a DataFrame out, creating missing folders
preview_data	Shape, dtypes, missing counts and the first rows
summary_statistics	Descriptives with skew and kurtosis
correlation	Pearson, Spearman or Kendall

Estimation

Tool	Purpose
regression	OLS, optionally with HC or HAC standard errors
regression_diagnostics	Breusch-Godfrey, White and Jarque-Bera in one call
unit_root	ADF or KPSS, differencing until stationary

Charts

Tool	Purpose
plot	line, scatter, hist, box or bar — optionally straight to a file
save_figure	Write the open matplotlib figure to a file

Errors

Failures name the exception and the line of your code, and leave the session intact:

Python error: NameError on line 2: name 'undefined_name' is not defined

Anything printed before the failure is reported with it, since that output is often what explains the failure.

Figures

A plot cannot come back as text. Draw it, then save it:

py.run("import matplotlib; matplotlib.use('Agg')")
py.run("import matplotlib.pyplot as plt; plt.plot(df['x'], df['y'], 'o')")
py.save_figure("figures/scatter.png")     # missing folders are created

Worth knowing

• The interpreter is separate from the server's. Check python_status before relying on a package being there.
• State is a convenience and a hazard. Names persist, so a stale variable from an earlier step can quietly feed a later one. reset_namespace when starting something new.
• run_python executes whatever it is given, in your environment, with your file access. That is the point of it, and worth being deliberate about.
• The final expression is echoed. A long DataFrame will print in full unless you slice it.

Tests

python tests/test_live.py     # 41 tests

They cover persistence, error reporting with line numbers, JSON round trips, figure writing, recovery after user code kills the worker outright, and every data and estimation tool against generated data with known coefficients.

Checked against EViews on the same data, the regression agrees to every printed digit -- coefficients, R-squared, Durbin-Watson, and the Breusch-Godfrey and Jarque-Bera statistics alike.

Licence

MIT. Copyright (c) 2026 Merwan Roudane.