Pixi¶

a faster, reproducible replacement for conda/mamba

Pixi is a package manager built on the conda ecosystem — it resolves packages from the same channels (conda-forge, bioconda) and produces fully compatible environments. It is written in Rust, uses a significantly faster solver, and is designed around project-level reproducibility from the start.

Why switch?

If you already use conda or mamba, the learning curve is minimal. The payoff is faster installs, exact reproducibility, and no base environment cluttering your shell.

Loading Pixi on BMRC¶

Pixi is available as a module — there is no base environment to activate, no conda init, and no changes to your .bashrc:

module load Pixi

Each project carries its own isolated environment stored in a .pixi/ directory inside the project folder.

Getting started¶

1. Initialise a project¶

In your project directory:

pixi init .

This creates a pixi.toml file that describes your project's dependencies, channels, and platforms. It is the equivalent of environment.yml, but machine-written and kept in sync automatically.

2. Add packages¶

# From conda-forge (default)
pixi add numpy scipy scanpy

# From bioconda
pixi add --channel bioconda samtools

Every pixi add resolves the full dependency graph and updates pixi.lock — a file that records the exact URL and hash of every package for every platform.

3. Run your code¶

Rather than activating an environment, use pixi run:

pixi run python my_analysis.py
pixi run jupyter lab

For interactive sessions (e.g. debugging, exploratory work), you can drop into a shell with the environment active:

pixi shell

4. Commit your lock file to Git¶

git add pixi.toml pixi.lock
git commit -m "Add pixi environment"

Anyone cloning the repository can then reproduce the exact environment with:

pixi install

Using Pixi in Slurm job scripts¶

Replace conda activate with pixi run in your job scripts:

conda / mambaPixi

#!/bin/bash
#SBATCH ...
source activate myenv
python my_analysis.py

#!/bin/bash
#SBATCH ...
module load Pixi
pixi run python my_analysis.py

If your script runs many commands and you want to avoid prefixing each one with pixi run, use the shell hook to activate the environment in-place:

module load Pixi
eval "$(pixi shell-hook)"

python step_one.py
Rscript step_two.R

PyPI packages¶

Pixi has native PyPI support alongside conda packages — no separate pip install calls needed. Add a [pypi-dependencies] section to your pixi.toml:

[pypi-dependencies]
my-package = ">=1.2.0"

Or via the command line:

pixi add --pypi my-package

Pixi resolves conda and PyPI dependencies together, avoiding the silent conflicts that can arise from mixing conda install and pip install.

Command reference¶

Task	conda / mamba	Pixi
Install package	`conda install numpy`	`pixi add numpy`
Install from bioconda	`conda install -c bioconda samtools`	`pixi add --channel bioconda samtools`
Run command in environment	`conda activate myenv && python ...`	`pixi run python ...`
Interactive shell	`conda activate myenv`	`pixi shell`
List installed packages	`conda list`	`pixi list`
Reproduce environment	`conda env create -f environment.yml`	`pixi install`
Update all packages	`conda update --all`	`pixi update`
Remove environment	`conda env remove -n myenv`	`rm -rf .pixi/`

Migrating from conda¶

From an `environment.yml`¶

If you have an existing conda environment file, Pixi can import it directly:

pixi init --import environment.yml

This creates a pixi.toml from your channels, dependencies, and constraints, then generates a fresh pixi.lock. Review the output to confirm everything looks correct before removing the old environment.

From an active conda environment¶

If you do not have an environment.yml but have an active environment, export it first:

conda activate myenv
conda env export --from-history > environment.yml
pixi init --import environment.yml

--from-history

This exports only the packages you explicitly requested, not the full transitive dependency tree. Pixi will re-resolve dependencies from scratch, which gives a cleaner result than exporting the entire pinned environment.

Verifying the migration¶

Once imported, check that everything resolves as expected:

pixi install
pixi run python -c "import scanpy; print(scanpy.__version__)"

Why Pixi is faster and more reproducible¶

Speed. Benchmarked on a BMRC head node, creating a scanpy + anndata environment:

Tool	Command	Wall time
mamba	`mamba create -n bench_env -c conda-forge -c bioconda scanpy anndata -y`	1 min 39 s
Pixi	`pixi add scanpy scipy`	14.9 s

Pixi resolved and installed the same packages ×6.7 faster than mamba — and mamba is already the fastest conda-compatible solver available.

Reproducibility. Conda environments created from environment.yml are not fully reproducible: running conda env create six months later can silently install different transitive dependency versions as packages on conda-forge update. The --explicit workaround produces a platform-specific file that cannot be shared between Linux and macOS.

Pixi solves this with pixi.lock, which records the exact URL and hash of every package, for every platform, in a single human-readable file:

	conda + `environment.yml`	Pixi + `pixi.lock`
Exact reproducibility	No	Yes
Cross-platform lock file	No (`--explicit` is Linux-only)	Yes
VCS-friendly	Partial	Yes
Auto-updated on changes	No	Yes