Register a jupyter kernel for the current pyenv (with fix for Jupyter servers running in virtualenvs).

register-jupyter-env

#!/bin/sh

if [ "$PYENV_VERSION" -ne "" ]
then
    name=`pyenv version-name`
    python=`pyenv which python`
else
    name=`basename "$VIRTUAL_ENV"`
    python="$VIRTUALENV/bin/python"
fi
jupyterdir=$(jupyter --data-dir)
kerneldir="${jupyterdir}/kernels/${name}"

echo "Installing jupyter kernel file $name for $python to $kerneldir ..."

pip install ipykernel
mkdir -p "$kerneldir"
cat > "$kerneldir"/kernel.json <<EOF
{
    "argv": [ "$python", "-m", "ipykernel", "-f", "{connection_file}" ],
    "display_name": "$name", 
    "language": "python",
    "env": {"__PYVENV_LAUNCHER__": ""}
}
EOF

cat "$kerneldir"/kernel.json 

Following @thvitt's general approach, I was eventually able to set up completely isolated Jupyter kernels for multiple different Pythons (cpy37, cpy38, cpy39, etc) installed using pyenv.

However, becuase I run my Jupyter servers from within virtual environments (created using the builtin python -m venv command), I ran into some challenging bugs along the way. The fix was to add an env entry to kernel.json that zeros out the __PYVENV_LAUNCHER__ variable.

The __PYVENV_LAUNCHER__ bug was also fixed upstream in cpython itself in 3.9.0, and backported to 3.7.9 and 3.8.? (not sure exactly which minor version). So if you run into this bug, you can also fix it by just upgrading your virtual env to the latest Python release.