GitHub - elohmeier/nixwrap: Wrap Nix binaries into Python packages for installation via uv tool install

3 min read Original article ↗

Run Nix packages from Python without Nix installed.

Overview

nixwrap provides two ways to run Nix-built binaries:

  1. Direct CLI: Run any package from nixpkgs instantly
  2. Python packages: Install tools via pip/uv from a PEP 503 index

Both methods fetch binaries from the Nix binary cache, patch them for relocation, and run them using a bundled dynamic linker.

Quick Start

CLI (Recommended)

# Install nixwrap
pip install nixwrap

# Run any nixpkgs package directly
nixwrap ripgrep --version
nixwrap jq --help
nixwrap fd --version

# Or use uvx for one-off execution
uvx nixwrap bat README.md

Python Package Index

# One-off execution with uvx
uvx --index-url https://elohmeier.github.io/nixwrap fd --version

# Or via pipx
uvx --with pip pipx run --index-url https://elohmeier.github.io/nixwrap fd --version

# Install a tool permanently
uv tool install --index-url https://elohmeier.github.io/nixwrap ripgrep

# Use the tool
rg --version

How It Works

Both modes share the same core process:

  1. Package Discovery: Queries the nix-index-database to find packages and their binaries
  2. Closure Computation: Fetches narinfo files to compute the full dependency closure
  3. Binary Fetching: Downloads and extracts NAR archives from the Nix binary cache
  4. Patching: Uses patchelf to fix binaries with hardcoded /nix/store paths
  5. Execution: Runs binaries via a bundled ld-linux with the correct library path

CLI Mode

When using nixwrap <package>, binaries are cached locally:

~/.cache/nixwrap/
  nix/store/          # Extracted packages and dependencies
  packages/           # Package metadata cache (speeds up repeated runs)

The first run downloads and patches the binary. Subsequent runs use the cached version.

Index Mode

When installing from the PEP 503 index, the build backend:

  1. Fetches all dependencies at wheel build time
  2. Patches binaries and bundles them directly into the wheel
  3. Installs the wheel with all binaries as package data

This results in a self-contained Python package with no external cache.

Repository Structure

nixwrap/
  pyproject.toml              # Main package configuration
  src/nixwrap/
    __init__.py
    backend.py                # PEP 517 build backend for wheels
    cli.py                    # CLI entry point (nixwrap command)
    index.py                  # nix-index-database parser
    patcher.py                # ELF patching with patchelf
  nixwrap-index/              # Separate package containing nix-index data
    pyproject.toml
    src/nixwrap_index/
      __init__.py
      data/                   # Index files (downloaded in CI)
  tools/
    generate_stubs.py         # Generates stub sdists for PEP 503 index
  .github/workflows/
    publish-pages.yml         # Publishes stubs to GitHub Pages
    publish-pypi.yml          # Publishes nixwrap to PyPI
    publish-pypi-index.yml    # Publishes nixwrap-index to PyPI

Requirements

  • Python 3.14+ (for compression.zstd stdlib module)
  • Linux x86_64 or aarch64

Development

# Clone and setup
git clone https://github.com/elohmeier/nixwrap.git
cd nixwrap
uv sync

# Run the CLI locally
uv run nixwrap ripgrep --version

# Build wheels locally
uv build

Limitations

  • Hardcoded runtime paths: Some packages have /nix/store/... paths compiled into the binary for runtime files (config, data, plugins). These paths can't be patched with patchelf. Examples:

    • neovim-unwrapped: Works for --version but fails at runtime because it can't find Lua modules and syntax files at the hardcoded paths.
    • Simple CLI tools like ripgrep, fd, jq, bat work fine since they don't depend on external runtime files.
  • Wrapper scripts: Packages that use shell wrapper scripts (like neovim, python) won't work because they have hardcoded paths in bash scripts, not ELF binaries.

  • Linux only: Currently only supports Linux (x86_64 and aarch64).

  • Python 3.14+: Requires Python 3.14 for the stdlib zstd module.

License

MIT