python bond

Ambivalent bonds between Python and other languages

The Python module bond supports transparent remote/recursive evaluation between Python and another interpreter through automatic call serialization.

In poorer words, a bond lets you call functions in other languages as they were normal Python functions. It also allows other languages to call Python functions as if they were native.

Remote output is also transparently redirected locally, and since the evaluation is performed through a persistent co-process, you can actually spawn interpreters on different hosts through “ssh” efficiently.

bond currently supports PHP, Perl, JavaScript (Node.js) and Python itself.


This module is available directly on the Python Package Index.

Latest 5 available releases of python-bond (most recent first):

Full archive in releases/.

Files as found in the latest release of python-bond:

NEWS:Summary of changes between releases.

Release announcements are made on the mailing list.

All the relevant source/developer information can be found on Gitlab:


>>> # Let's bond with a PHP interpreter
>>> from bond import make_bond
>>> php = make_bond('PHP')
>>> php.eval_block('echo "Hello world!\n";')
Hello world!

>>> # Make an expensive split function using PHP's explode
>>> split = php.callable('explode')
>>> split(' ', "Hello world splitted by PHP!")
[u'Hello', u'world', u'splitted', u'by', u'PHP!']

>>> # Call Python from PHP
>>> def call_me():
...     print("Hi, this is Python talking!")
>>> php.export(call_me)
>>> php.eval('call_me()')
Hi, this is Python talking!

>>> # Use some remote resources
>>> remote_php = make_bond('PHP', 'ssh remote php')
>>> remote_php.eval_block('function call_me() { echo "Hi from " . system("hostname") . "!"; }')
>>> remote_php.eval('call_me()')
Hi from remote!

>>> # Bridge two worlds!
>>> perl = make_bond('Perl')
>>> php.proxy('explode', perl)
>>> # note: explode is now available to Perl, but still executes in PHP
>>> perl.eval('explode("=", "Mind=blown!")')
[u'Mind', u'blown!']

Practical examples

Incremental code-base migration

I originally needed bond for migrating a large PHP project to Python. With bond you can rewrite a program incrementally, while still executing all the existing code unchanged. You can start by rewriting just a single function in an empty shell, wrapping your existing code:

from bond import make_bond
import sys

php = make_bond('PHP')

def new_function(arg)
    # do something here

php.export(new_function, 'function_to_be_replaced')'main', sys.argv)

Mixing Python 2/3 code bases

You can use bond to mix Python 2/3 code. Python <=> Python bonds automatically use pickling as a protocol, which makes serialization almost invisible.

In this scenario, you can start writing new code directly on Python 3, while using Python 2 only for the libraries which are still missing.

For example, you can use Mechanize on Python 3 with minimal changes:

from bond import make_bond
py2 = make_bond('Python', 'python2', trans_except=False)
py2.eval_block('import mechanize; br = mechanize.Browser()')'', '')
title ='br.title')

eval_block is only being used as an example here to make it self-contained. A more reasonable solution for larger chunks of code is to split the source into a distinct file that can be loaded at once in the remote interpreter:

from bond import make_bond
py2 = make_bond('Python', 'python2', trans_except=False)
py2.eval_block('import .mypython2lib')

This reduces the amount of clutter and keeps the distinction between new and legacy code clear. You should also keep in mind that since the remote language is itself Python, expressions themselves (for whenever call is insufficient) can be quoted just by using repr.

Similarly, you can use bond to combine regular CPython and PyPy runtimes (all the required modules work as expected on PyPy):

from bond import make_bond
pypy = make_bond('Python', 'pypy')

Remote/parallel computation

You can easily use bond to perform remote/parallel computation. Nobody stops you from having multiple interpreters at the same time: you can create multiple bonds to setup a poor-man’s distributed system with minimal effort:

# setup the workers
from bond import make_bond
hosts = ['host1', 'host2', 'host3']
nodes = [make_bond('Python', 'ssh {} python'.format(host)) for host in hosts]

# load our libraries first
for node in nodes:
    node.eval_block('from library import *')

# execute "do_something" remotely on each worker
from threading import Thread
threads = [Thread(target=lambda:'do_something')) for node in nodes]
for thread in threads: thread.start()

# collect the results
results = [thread.join() for thread in threads]

Distributed producer/consumer schemes also come for free by proxying calls:

host1.eval_block(r'''def consumer(data):
   # do something with data

host2.eval_block(r'''def producer():
    while True:
       data = function()

host1.proxy('consumer', host2)'producer')

It’s even more interesting if you realize that the producers/consumers don’t even have to be written in the same language, and don’t know that the call is actually being forwarded.

bond doesn’t even need to be installed remotely: the required setup is injected directly into a live interpreter. The wire protocol is simple enough that any language supporting an interactive REPL can be called. In fact, the drivers themselves are designed to be used from any other language.

Privilege separation

There might be times when it makes sense to create an unprivileged context to perform potentially dangerous operations, such as decoding an uploaded file on which you have zero trust. A common approach would be to communicate with an unprivileged daemon built for the purpose, but it usually requires dedicated effort. Running such processes using bond instead is almost trivial:

# early in the setup phase of our daemon we create a bond using
# passwordless sudo, changing to an unprivileged user
py = make_bond('Python', 'sudo -u nobody python',
               trans_except=False, protocol='JSON')
py.eval_block('from mylibrary import decode_file')

# make decode_file() available as a normal function
decode_file = py.callable('decode_file')

# assuming decode_file() takes a file name which is at least readable by
# the unprivileged user, we can just take it's return value
data = decode_file(path)

Contrarily to other examples involving Python, here we actually restrict the serialization protocol to plain JSON. Nothing changes from the caller (our) perspective, except that the bond now can’t share with us anything beyond trivial types. Python <=> Python bonds use pickle by default, which is not sensible here as pickle allows arbitrary Python structures and handlers to be run (including bytecode itself).

If just running the context as another user is not enough, then setting up an LXC container doesn’t add much complexity, since we can just use lxc-execute to attach directly to the new instance’s STDIO:

py = make_bond('Python', 'lxc-execute -n <name> -f <config> /path/to/python',
               trans_except=False, protocol='JSON')

This way an ephemeral container is started/destroyed automatically along with our daemon. The container itself can expose just a few shared/read-only directories, or nothing at all if the entire I/O is built on top of bond.



A bond.Bond object is not normally constructed directly, but by using the bond.make_bond() function:

import bond
interpreter = bond.make_bond('language')

The first argument should be the desired language name (“JavaScript”, “PHP”, “Perl”, “Python”). The list of supported languages can be fetched dynamically using bond.list_drivers().

You can override the default interpreter command using the second argument, which allows to specify any shell command to be executed:

import bond
py = bond.make_bond('Python', 'ssh remote python3')

An additional list of arguments to the interpreter can be provided using the third argument, args:

import bond
py = bond.make_bond('Python', 'ssh remote python3', ['-E', '-OO'])

The arguments, contrarily to the command, are automatically quoted.

Some command line arguments may be supplied automatically by the driver to force an interactive shell; for example “-i” is supplied if Python is requested. You can disable default arguments by using def_args=False.

The following keyword arguments are supported:


Working directory for the interpreter (defaults to current working directory).


Environment for the interpreter (defaults to os.environ).


Enable (default) or suppress default, extra command-line arguments to the interpreter.


Defines the timeout for the underlying communication protocol. Note that bond cannot distinguish between a slow call or noise generated while the interpreter is set up. Defaults to 60 seconds.


Accepts a file handle which is used to log the entire communication with the underlying interpreter for debugging purposes.


Enables/disables “transparent exceptions”. Exceptions are always first class, but when trans_except is enabled, the exception objects themselves will be forwarded across the bond. If trans_except is disabled (the default for all languages except Python), then local exceptions will always contain a string representation of the remote exception instead, which avoids serialization errors.


Forces a specific serialization protocol to be chosen. It’s automatically selected when not specified, and usually matches “JSON”.

bond.Bond Methods

The resulting bond.Bond class has the following methods:


Evaluate and return the value of a single statement of code in the top-level of the interpreter.


Execute a “code” block in the top-level of the interpreter. Any construct which is legal by the current interpreter is allowed. Nothing is returned.


Return a reference to an single, unevaluated statement of code, which can be later used in eval(), eval_block() or as an immediate argument to call(). See Quoted expressions.


Terminate the communication with the interpreter.

call(name, *args):

Call a function “name” in the interpreter using the supplied list of arguments *args (apply *args to a callable statement defined by “name”). The arguments are automatically converted to their other language’s counterpart. The return value is captured and converted back to Python as well.


Return a function that calls “name”:

explode = php.callable('explode')
# Now you can call explode as a normal, local function
explode(' ', 'Hello world')

export(func, name):

Export a local function “func” so that can be called on the remote language as “name”. If “name” is not specified, use the local function name directly. Note that “func” must be a local function, not a function name.

proxy(name, other, remote):

Export a function “name” from the current bond to “other”, named as “remote”. If “remote” is not provided, the same value as “name” is used.


Start an interactive session with the underlying interpreter. By default, all input lines are executed with bond.eval_block(). If “!” is pre-pended, execute a single statement with bond.eval() and print it’s return value. You can continue the statement on multiple lines by leaving a trailing “\”. Type Ctrl+C to abort a multi-line block without executing it.


All exceptions thrown by the bond module are of base type RuntimeError <= BondException.

Thrown during initialization or unrecoverable errors.
Thrown when the bond exits unexpectedly.
Thrown when an object/exception which is sent or received cannot be serialized by the current protocol. The side attribute can be either “local” (when attempting to send) or “remote” (when receiving). A SerializationException is not fatal.
Thrown for uncaught remote exceptions. The “data” attribute contains either the error message (with trans_except=False) or the remote exception itself (trans_except=True).

Beware that both SerializationException (with side="remote") and RemoteException may actually be originating from uncaught local exceptions when an exported function is called. Pay attention to the error text/data in these cases, as it will contain several nested exceptions.

Quoted expressions

bond has minimal support for working with quoted expressions, through the use of Bond.ref(). ref() returns a reference to a unevaluated statement that can be fed back to eval(), eval_block(), or as an immediate (i.e.: not nested) argument to call(). References are bound to the interpreter that created them.

ref() allows to “call” methods that take remote un-serializable arguments, such as file descriptors, without the use of a support function and/or eval:

pl = make_bond('Perl')
pl.eval_block('open($fd, ">file.txt");')
fd = pl.ref('$fd')'syswrite', fd, "Hello world!")'close', fd)

Since ref() objects cannot be nested, there are still cases where it might be necessary to use a support function. To demonstrate, we rewrite the above example without quoted expressions, while still allowing an argument (“Hello world!”) to be local:

pl = make_bond('Perl')
pl.eval_block('open($fd, ">file.txt");')
pl.eval_block('sub syswrite_fd { syswrite($fd, shift()); };')'syswrite_fd', "Hello world!")

Or more succinctly:'sub { syswrite($fd, shift()); }', "Hello world!")

Language support


Python, as the identity language, has no restriction on data types. Everything is pickled on both sides, including exceptions.


  • Performed locally and remotely using cPickle in Python 2 or pickle in Python 3.
  • Serialization exceptions on the remote side are of base type TypeError <= _BOND_SerializationException.

Python 2 / Python 3 / PyPy:

You can freely mix Python versions between hosts/interpreters (that is: you can run Python 3 code from a Python 2 host and vice-versa). You’ll need to disable transparent exceptions between major versions though, as the exception hierarchy is different:

# assuming a python2.7 environment
from bond import make_bond
py = make_bond('Python', 'python3', trans_except=False)



  • The PHP’s >= 5.3 command line interpreter needs to be installed. On Debian/Ubuntu, the required package is php5-cli.


  • Performed remotely using JSON. Implement the JsonSerializable interface to tweak which/how objects are encoded.
  • Serialization exceptions on the remote side are of base type _BOND_SerializationException. The detailed results of the error can also be retrieved using json_last_error.


  • PHP <= 5.3 doesn’t support the JsonSerializable interface, and thus lacks the ability of serializing arbitrary objects.
  • You cannot use call on a built-in function such as “echo”. You have to use a real function instead, like “print”. You can still call “echo” by using eval or eval_block.
  • Unfortunately, you cannot catch “fatal errors” in PHP. If the evaluated code triggers a fatal error it will terminate the bond without appeal. A common example of such error can be attempting to use an undefined variable or function (which could happen while prototyping).
  • Due to the inability to override built-in functions, error_reporting() is not completely transparent and always returns 0. It shouldn’t be used to control the display error level. Use _BOND_error_reporting() instead, which has the same usage/signature as the built-in function.


Perl is a quirky language, due to its syntax. We assume here you’re an experienced Perl developer.


  • Perl >= 5.14 is required, with the following modules:

    • JSON
    • Data::Dump
    • IO::String

    On Debian/Ubuntu, the required packages are libjson-perl libdata-dump-perl and libio-string-perl.


  • Performed remotely using JSON. Implement the TO_JSON method on blessed references to tweak which/how objects are encoded.
  • Serialization exceptions on the remote side are generated by dying with a _BOND_SerializationException @ISA.


  • By default, evaluation is forced in array context, as otherwise most of the built-ins working with arrays would return an useless scalar. Use the “scalar” keyword for the rare cases when you really need it to.

  • You can “call” any function-like statement, as long as the last argument is expected to be an argument list. This allows you to call builtins directly:'map { $_ + 1 }', [1, 2, 3])
  • You can of course “call” a statement that returns any CODE. Meaning that you can call references to functions as long as you dereference them first:'&$fun_ref', ...)'&{ $any->{expression} }', ...)

    Likewise you can “call” objects methods directly:'$object->method', ...)
  • eval_block introduces a new block. Variables declared as “my” will not be visible into a subsequent eval_block. Use a fully qualified name or “our” to define variables that should persist across blocks:

    perl.eval_block('our $variable = 1;')


JavaScript is supported through Node.js.


  • Node.js v0.6.12 and v0.10.29 have been tested. On Debian/Ubuntu, the required package is nodejs.


  • Performed remotely using JSON. Implement the toJSON property to tweak which/how objects are encoded.
  • Serialization exceptions on the remote side are of base type TypeError <= _BOND_SerializationException.


  • Currently the code expects an unix-like environment with /dev/stdin to perform synchronous I/O.

  • Since there’s no distinction between “plain” objects (dictionaries) and any other object, almost everything will be silently serialized. Define a custom “toJSON” property on your “real” objects to control this behavior.

  • When executing a remote JavaScript bond with Node.js <= 0.6, you need to manually invoke the REPL, as follows:

    js = make_bond('JavaScript',
                   "ssh remote node -e 'require\(\\\"repl\\\"\).start\(\)'",

    When executing “node” locally, or when using Node.js >= 0.10, this is not required (the “-i” flag is automatically provided).

Common limitations

  • Except for Python, only basic types (booleans, numbers, strings, lists/arrays and maps/dictionaries) can be transferred between the interpreters.
  • Serialization is performed locally using JSON. Implement a custom JSONEncoder to tweak which/how objects are encoded.
  • If an object that cannot be serialized reaches a “call”, “eval”, or even a non-local return such as an error or exception, it will generate a SerializationException on the local (Python) side.
  • Strings are always UTF-8 encoded.
  • References are implicitly broken as objects are transferred by value. This is obvious, as you’re talking with a separate process, but it can easily be forgotten due to the blurring of the boundary.
  • Calling functions across the bridge is slow, also in Python, due to the serialization. But the execution speed of the functions themselves is not affected. This might be perfectly reasonable if there are only occasional calls between languages, and/or the calls themselves take a significant fraction of time.

General/support mailing list

If you are interested in announcements and development discussions about bond, you can subscribe to the bond-devel mailing list by sending an empty email to <>.

You can contact the main author directly at <>, though using the general list is encouraged.