Interact with python's numpy package from the command line. Useful as part of pipelines.
Switch branches/tags
Nothing to show
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Type Name Latest commit message Commit time
Failed to load latest commit information.

Travis CI


Interact with python's numpy package from the command line. Useful as part of pipelines.


Influenced by, and liberally taking ideas from from Wes Turner's pyline utility.


For convenience certain features provide interfaces to pandas and matplotlib. These are not installed by default to minimize the number of dependencies for basic usage. Tested with python 2.7 and python 3.5.


Command line pipelines are wonderful things. Some nice properties they have include:

  • Complete searchable history of everything you have run
  • Being able to use shell commands you already know rather than learning new python libraries
  • Being able to compose disparate commands through string input and output
  • Completion

However the command line is sometimes slightly... lacking. Particularly when it comes to things like maths. There are ad-hoc, single purpose commands that can help: things like feedgnuplot or sum or similar, but they will always only solve one problem.

Here we try to solve a general class of problems by welding python (any numpy) to the command line. This means that anything you can do in python can be done in a way that easily interacts with a the command line.


# The squares of the numbers 1 to 100
seq 100 | npcli 'd**2'

# Work out the mean of some random numebrs
npcli 'np.random.random(10000)' -m numpy.random | npcli 'np.mean(d)'

# Plot a graph
seq 100 | npcli -nK 'pylab.plot(d);'

# Produce a histogram of when most lines in syslog are printed
sudo cat /var/log/syslog | cut -d " "  -f 1-4 | xargs -L 1 -I A date -d  A  +%s | npcli 'd % 86400' | npcli 'd // 3600 * 3600' | uniq -c  | npcli -Kn 'pylab.plot(d[:,1], d[:,0]);'

# Generate some random data
npcli -K 'random(100)'

# Summarize the last 100 days of GOOG's share price
curl "" | head -n 100 | npcli  -I pandas  'd["Close"].describe()' -D

# Chain together operations
seq 10 | npcli  'd' -e 'd*2' -e 'd + 4' -e 'd * 3' -e 'd - 12'  -e 'd / 6'

# Multiple data sources
npcli --name one <(seq 100) --name two <(seq 201 300) 'one + two'


usage: npcli [-h] [--expr EXPR] [--code] [--debug]
             [--input-format INPUT_FORMAT] [--kitchen-sink] [--name NAME NAME]
             [--output-format OUTPUT_FORMAT | --raw | --repr | --no-result]
             [--module MODULE] [-f data_source]
             expr [data_sources [data_sources ...]]

Interact with numpy from the command line

positional arguments:
  expr                  Expression involving d, a numpy array
  data_sources          Files to read data from. Stored in d1, d2 etc

optional arguments:
  -h, --help            show this help message and exit
  --expr EXPR, -e EXPR  Expression involving d, a numpy array. Multipe
                        expressions get chained
  --code                Produce python code rather than running
  --debug               Print debug output
  --input-format INPUT_FORMAT, -I INPUT_FORMAT
                        Dtype of the data read in. "lines" for a list of
                        lines. "str" for a string. "csv" for csv, "pandas" for
                        a pandas csv, "json" for json data
  --kitchen-sink, -K    Import a lot of useful things into the execution scope
                        A named data source
                        Output as a flat numpy array with this format. "str"
                        for a string
  --raw                 Result is a string that should be written to standard
  --repr, -D            Output a repr of the result. Often used for _D_ebug
  --no-result, -n       Discard result
  --module MODULE, -m MODULE
                        Result is a string that should be written to standard
  -f data_source


Bleeding edge

pip install git+

Stable release

pip install git+git://

Alternatives and prior work


There are unit tests: you can run them with

python test

This will run the tests using tox, testing the installation commands and different versions of python.

For a quicker test run use:

nosetests test


npcli uses argparse. argparse appears to be not be able to deal with repeated flags (-e 1 -e second) and repeated optional position args (i.e. data sources), it may error out when given valid input. This can be circumvented by using the -f flag in preference to positional arguments. However, we still allow positional arguments in the interest of discoverability. I'm open to this being a bad decision.

Just open a file for goodness sake

It is very easy to do more on the command line than one should. Everything that is done here can be done in a python file with calls to subprocess. Above a certain size, one-liners become unwieldy

The cost of scripting in python is that you actually have to go to the effort of opening file, and doing the kind of things npcli automates can take quite a lot of boilerplate.

One also loses the simplicity of the shell debug cycle: "modify", "press enter", "see if it works".