↓ Skip to main content

The toolchain behind everything here

One file. Every tool. Any machine.

Every project I build runs on conda-forge and Pixi from prefix.dev. This blog included: the web server, the site generator, Python, Node, compilers and the AI tooling all come from one Pixi file. No sudo, no system packages, no "works on my machine". Here's why I'm such a fan.

Two parts, one experience

conda-forge: the packages

A community-run collection of open-source packages, built for Linux, macOS and Windows on several CPU architectures. Not just Python: C and C++ libraries, compilers, databases, web servers, Go, Rust, Node and command-line tools.

Pixi: the workflow

A fast package and project manager written in Rust by prefix.dev. One pixi.toml declares your tools, a lock file pins them exactly, and tasks replace your Makefile and README instructions. PyPI packages work too.

Together: it just works

Clone a project, run one command, and you have the exact environment I had, on your laptop, a server or a CI runner. Nothing touches the rest of your system.

39

tasks run this blog, from one Pixi file.

Building the site, deploying it, generating these slide decks and PDFs, the weekly email, LinkedIn reminders, the Ask the blog search index, traffic reports. Each is pixi run <task>, with the right tools already in place.

This site, for real

What a Pixi file looks like in practice

[workspace]
channels = ["conda-forge"]
platforms = ["linux-aarch64"]

[dependencies] nginx = “>=1.29.6,<2” # the web server hugo = “>=0.157.0,<0.158” # the site generator nodejs = “>=25.8.1,<25.9” go = “>=1.26.1,<2” weasyprint = “>=67.0,<68” # PDFs for subscribers c-compiler = “>=1.11.0,<2”

[tasks] deploy = “… build images, bonuses and slides, run hugo, reload nginx” build-slides = “python conf/build-slides.py”

An excerpt of the real file. One pixi install on a fresh server gives me everything, at the exact versions in the lock file.

Why it matters to a security person

Reproducible, pinned, and nothing installed as root.

Every dependency is declared and locked, so I know exactly what runs. Environments live in the project folder, not in the operating system, so there's nothing to install with sudo and nothing to clean up. That's a big part of a sane software supply chain.

Real moments where it saved me

How I use it, day to day

  1. Patching a zero-day the same nightWhen the NGINX2Root vulnerabilities hit, the fixed nginx wasn't packaged yet. pixi exec gave me a C compiler on the spot, and I built the patched release from source against the libraries already in my Pixi environment.
  2. Throwaway tools without installing anythingpixi exec -s poppler pdftoppm ... to render a PDF, used once, gone. No system packages, no leftovers.
  3. Several environments in one projectThe blog has a default environment and a separate one for notebooks, each with its own dependencies, side by side.
  4. Shipping environments to other machinesRixi sends a project and its exact Pixi environment to a remote box and runs it there.
  5. A whole data lake for under 10 eurosDuckLake on Hetzner with Pixi: catalog, storage and access control, reproducible from one file.
  6. Building and packaging native codeTurboQuant with llama.cpp: Pixi builds and conda packages for GPU inference on consumer hardware.

Try it in two minutes

Get started

curl -fsSL https://pixi.sh/install.sh | sh   # install Pixi (see pixi.sh for Windows)
pixi init my-project && cd my-project
pixi add python duckdb                         # packages from conda-forge
pixi add --pypi requests                       # PyPI works too
pixi task add hello "python -c 'print(42)'"
pixi run hello

Read more on supply chain security with Pixi and conda in A Layered Architecture for Python and Conda Supply Chain Security.

Thank you, conda-forge and prefix.dev.

conda-forge is run by volunteers and maintainers around the world, and Pixi is built in the open by the team at prefix.dev. If you use them, contribute a recipe, report an issue, or just tell them it helped.

My personal view as a happy user. Not sponsored by, or affiliated with, prefix.dev or conda-forge.