mise-lang-python-packages

Python package managers — the 2026 answer

Python's package management has been a mess for a decade. The good news: uv (by Astral, same folks as ruff) has become the clear winner for new projects, and the rules are finally simple.

The layering

mise           pins the Python interpreter
 └─> uv        manages the project's venv and lockfile
      └─> installed packages in .venv/

And separately, for global CLIs:

mise
 └─> pipx:<tool>    installs global Python CLIs in isolated venvs

Rule: mise handles "which Python". uv handles "which packages, in this project". pipx handles "which Python CLIs, globally". Don't cross the streams.

The pick — 2026 defaults

You want to…	Use
Start a new Python project	uv
Join a poetry-using project	poetry (don't migrate for no reason)
Install a global Python CLI (ruff, black, httpie, jupyterlab)	pipx (via mise's pipx: backend)
Run a single script with deps	uv (uv run --with requests script.py)
Rails-style "just make it work" for a small script	uv
Scientific computing with conda-style env	uv (supports requirements.txt style) or pixi (if you need bioconda etc.)

The 2026 default is uv. It replaces pip, pip-tools, virtualenv, pyenv (no, mise), pipx (partially), and poetry. It's 10-100x faster than pip for installs. The project lock format is stable.

uv + mise — the canonical setup

[tools]
python = "3.12"
"pipx:uv" = "latest"   # install uv itself via mise

[env]
MISE_PYTHON_COMPILE = "0"
_.python.venv = { path = ".venv", create = true }

[tasks.install]
run = "uv sync"
sources = ["pyproject.toml", "uv.lock"]

[tasks.test]
depends = ["install"]
run = "uv run pytest"

[tasks.lint]
depends = ["install"]
run = "uv run ruff check ."

[tasks.fmt]
depends = ["install"]
run = "uv run ruff format ."

pyproject.toml:

[project]
name = "myapp"
version = "0.1.0"
requires-python = ">=3.12"
dependencies = [
  "fastapi>=0.110",
  "pydantic>=2.0",
]

[dependency-groups]
dev = [
  "pytest>=8.0",
  "ruff>=0.6",
]

[build-system]
requires = ["hatchling"]
build-backend = "hatchling.build"

uv sync reads this, creates/updates .venv, writes uv.lock. uv run pytest executes pytest inside the venv. Clean, fast, deterministic.

poetry + mise

If you're joining a poetry project, don't migrate to uv just for the sake of it. Poetry works fine with mise:

[tools]
python = "3.12"
"pipx:poetry" = "1.8"

[env]
MISE_PYTHON_COMPILE = "0"

[tasks.install]
run = "poetry install"

[tasks.test]
run = "poetry run pytest"

Poetry manages its own venv (usually in ~/Library/Caches/pypoetry/virtualenvs/ or .venv if you set poetry config virtualenvs.in-project true). mise pins the Python that poetry uses to create that venv.

Don't use both _.python.venv and poetry — poetry creates its own venv and gets confused.

pipx via mise — for global CLIs

[tools]
python = "3.12"
"pipx:ruff" = "latest"
"pipx:httpie" = "latest"
"pipx:jupyterlab" = "latest"
"pipx:black" = "24.10"

Each pipx: entry installs the package into its own isolated venv under mise's install dir. The CLI is shimmed onto PATH. Same interface as a normal pipx install, but version-pinned and mise-lock'd.

Use pipx for: global tools (ruff, black, httpie, awscli, jupyterlab). Don't use pipx for: project dependencies (those go in pyproject.toml).

pip — the fallback

pip still exists and still works. Use it when:

• You're teaching Python to someone who's already confused.
• A tool's install instructions explicitly say pip install ... and you're in a venv.
• You're debugging a dep issue and want to try a specific version without bumping the lockfile.

Don't use pip install globally (outside a venv). macOS and Ubuntu now refuse by default (PEP 668 "externally-managed-environment").

The mixing rules

1. Never mix poetry and uv in the same project. They fight over pyproject.toml semantics.
2. Never mix conda/mamba and mise Python. conda wants to own the Python install; mise wants to own the Python install. Pick one per project.
3. Never pip install into the mise-installed Python system site. Always use a venv.
4. Never sudo pip install anything, ever.
5. Don't have both uv.lock and poetry.lock in the same repo. Delete the unused one.

uv as a pyenv replacement?

uv can install and manage Python itself (uv python install 3.12). This overlaps with mise. Current guidance:

• Use mise for Python version pinning — it's the single version-manager for your whole stack (Python, Node, Go, etc.), so mise.toml is one place.
• Use uv for everything else Python — venv, deps, run.

Don't use uv python install when mise is already managing Python. They'd both try to own the install dir.

Scripting with uv

For single scripts, uv has a magic feature — inline dep metadata:

#!/usr/bin/env -S uv run --script
# /// script
# requires-python = ">=3.12"
# dependencies = ["httpx", "rich"]
# ///
import httpx
from rich import print
print(httpx.get("https://httpbin.org/ip").json())

uv run script.py creates an ephemeral venv with the listed deps and runs the script. Replaces half of the "just pip install requests" tutorials.

Anti-patterns

• Manually activating venvs via source .venv/bin/activate in mise tasks — use uv run / poetry run / _.python.venv auto-activation instead.
• pip install -r requirements.txt in a CI workflow with no lockfile — non-reproducible.
• Using pipx install and pip install --user interchangeably — pipx is strictly better for CLIs.
• Committing .venv/ — no. .gitignore it.
• Using conda inside a mise project — pick one.

See also

• mise-lang-python-overview — Python version resolution and compile flags.
• mise-migrate-from-pyenv — moving off pyenv.
• mise-env-directives — _.python.venv = { path, create } details.
• uv docs: docs.astral.sh/uv/.
• PEP 668: peps.python.org/pep-0668/.