envcore

pip freeze done right — only what your code actually uses.

pip freeze dumps your entire environment — often 200+ packages — most of which your project never touches. pipreqs scans source files but misses dynamic imports, conditional imports, and anything loaded at runtime.

Envcore hooks into Python's import system while your code actually runs, records only what gets imported, and writes a clean manifest. One command rebuilds that exact environment anywhere.

envcore trace train.py   →   env_manifest.json   →   envcore restore

No config files to maintain. No guesswork. Ship the manifest with your code.

Install

pip install envcore

Requires Python 3.9+.

Usage

Trace your script:

envcore trace train.py

  Traced 3 packages:

    numpy       1.26.4
    pandas      2.1.4
    torch       2.1.0

  Manifest saved → env_manifest.json

Restore anywhere:

envcore restore

That's it. The manifest contains only what was actually imported — nothing more.

Why runtime tracing?

Static tools read your source files. They miss:

• import torch inside an if cuda_available: block
• imports inside functions, called only at runtime
• dynamic importlib.import_module(...) calls
• packages imported by your dependencies on your behalf

Envcore catches all of these because it watches what Python actually loads, not what it might load.

Note: Envcore traces one execution. If your code has branches that aren't hit during tracing (e.g. a prod-only import), those won't be captured. Trace against a representative run.

Programmatic API

import envcore

envcore.start_tracking()

import numpy
import pandas
from sklearn.model_selection import train_test_split

envcore.stop_tracking()
envcore.snapshot("env_manifest.json")

Context manager:

from envcore import ImportTracer

with ImportTracer() as tracer:
    import numpy
    import pandas

print(tracer.module_names)  # ['numpy', 'pandas']

Decorator:

from envcore import ImportTracer

tracer = ImportTracer()

@tracer
def train():
    import torch
    import wandb

train()
print(tracer.module_names)  # ['torch', 'wandb']

Jupyter Notebooks

envcore notebook analysis.ipynb

Or inside a running notebook:

%load_ext envcore
%envcore start
import numpy, pandas
%envcore snapshot

How it works

1. TRACE    builtins.__import__ → ImportTracer hook
            Records module name, filters stdlib and pre-existing packages

2. RESOLVE  import name → PyPI package + pinned version
            "PIL" → Pillow 10.2.0
            "cv2" → opencv-python 4.9.0
            "sklearn" → scikit-learn 1.4.0

3. MANIFEST env_manifest.json
            { "numpy": "1.26.4", "torch": "2.1.0", ... }

4. RESTORE  pip install numpy==1.26.4 torch==2.1.0 ...

Other commands

Every command supports --help. Set NO_COLOR=1 to disable colour output.

Configuration

# pyproject.toml
[tool.envcore]
manifest = "env_manifest.json"
exclude = ["setuptools", "pip", "wheel"]
entry_points = ["src/main.py"]
auto_export = "requirements.txt"

Contributing

git clone https://github.com/JanBremec/envcore.git
cd envcore
pip install -e ".[dev]"
pytest tests/ -v

Issues and PRs welcome.

License

MIT — Jan Bremec