Eric Florenzano’s Blog

Writing an Markov-Chain IRC Bot with Twisted and Python

Nov 17, 2008

Twisted is one of Python's great secret weapons. It is an absolute workhorse, allowing for insanely fast network applications to be written with very little effort. So let's do what everyone does when they want to learn more about Twisted: let's write an IRC bot! This bot is going to use Markov Chains to simulate human speech. For whatever reason, I named this bot "YourMomDotCom".

First let's create a skeleton on top of which the rest of the bot will be created:

from twisted.words.protocols import irc
from twisted.internet import protocol

class MomBot(irc.IRCClient):
    def _get_nickname(self):
        return self.factory.nickname
    nickname = property(_get_nickname)

    def signedOn(self):
        print "Signed on as %s." % (self.nickname,)

    def joined(self, channel):
        print "Joined %s." % (channel,)

    def privmsg(self, user, channel, msg):
        print msg

class MomBotFactory(protocol.ClientFactory):
    protocol = MomBot

    def __init__(self, channel, nickname='YourMomDotCom'): = channel
        self.nickname = nickname

    def clientConnectionLost(self, connector, reason):
        print "Lost connection (%s), reconnecting." % (reason,)

    def clientConnectionFailed(self, connector, reason):
        print "Could not connect: %s" % (reason,)

We've now created an IRCClient subclass which will hold our application logic, and we've also written a factory class which will create instances of that MomBot client. Let's tie these together and start the event loop:

import sys
from twisted.internet import reactor

if __name__ == "__main__":
    chan = sys.argv[1]
    reactor.connectTCP('', 6667, MomBotFactory('#' + chan))

Now already we have a complete working IRC bot. Right now all it will do is connect to an IRC channel and echo all of the output to the command line. Not bad for how little code we've written. Now all we have to do is implement our application logic. Let's first start by creating the 'brain' of our Markov chain responder, and adding a function to train the brain:

from collections import defaultdict

markov = defaultdict(list)
STOP_WORD = "\n"

def add_to_brain(msg, chain_length, write_to_file=False):
    if write_to_file:
        f = open('training_text.txt', 'a')
        f.write(msg + '\n')
    buf = [STOP_WORD] * chain_length
    for word in msg.split():
        del buf[0]

In this, we are creating a defaultdict of lists. For every n-word sliding window, the word after that window is appended to the list of possible words. Here's an image which hopefully depicts better than words how the algorithm populates the brain:

But what good is a brain like this if we can't get words back from it. We're going to need to write a function to generate sentences from that brain:

def generate_sentence(msg, chain_length, max_words=10000):
    buf = msg.split()[:chain_length]
    if len(msg.split()) > chain_length:
        message = buf[:]
        message = []
        for i in xrange(chain_length):
    for i in xrange(max_words):
            next_word = random.choice(markov[tuple(buf)])
        except IndexError:
        if next_word == STOP_WORD:
        del buf[0]
    return ' '.join(message)

We start out our seed buffer with the first few words of the message, and if the message wasn't long enough, we fill the seed buffer with some random words from the markov's brain. Then we use the buffer as a key into the markov brain and randomly pick one of the values as our next word. Then we slide that buffer so that the chosen word is now the next word in the buffer (ejecting the oldest word in the buffer). If we ever see a stop word, we stop and return the generated sentence.

Now it's a matter of expanding our bot to take advantage of our markov brain. The first change we will need to make is to modify the MomBotFactory to take more parameters in its __init__ method:

class MomBotFactory(protocol.ClientFactory):
    protocol = MomBot

    def __init__(self, channel, nickname='YourMomDotCom', chain_length=3,
        chattiness=1.0, max_words=10000): = channel
        self.nickname = nickname
        self.chain_length = chain_length
        self.chattiness = chattiness
        self.max_words = max_words

    # ...

Now that it has all of that new information, we can add the final bit of functionality to MomBot responding using sentences generated from the markov brain by updating the privmsg method:

class MomBot(irc.IRCClient):

    # ...

    def privmsg(self, user, channel, msg):
        if not user:
        if self.nickname in msg:
            msg = re.compile(self.nickname + "[:,]* ?", re.I).sub('', msg)
            prefix = "%s: " % (user.split('!', 1)[0], )
            prefix = ''
        add_to_brain(msg, self.factory.chain_length, write_to_file=True)
        if prefix or random.random() <= self.factory.chattiness:
            sentence = generate_sentence(msg, self.factory.chain_length,
            if sentence:
                self.msg(, prefix + sentence)

If it gets a message from a non-user, we don't care about it, so we just disregard that message. Otherwise, if it's a message directed at the bot, we want to respond to the sending user so we add the user as a prefix to our next message. Then we grab a sentence from the markov brain and send it to the channel.

Not so bad. Now lets modify our startup portion of the script to re-train the brain with anything learned during the course of the bot's life. Here's how the new script will look:

if __name__ == "__main__":
        chan = sys.argv[1]
    except IndexError:
        print "Please specify a channel name."
        print "Example:"
        print "  python django-hotclub"
    if os.path.exists('training_text.txt'):
        f = open('training_text.txt', 'r')
        for line in f:
            add_to_brain(line, chain_length)
        print 'Brain Reloaded'
    reactor.connectTCP('', 6667, MomBotFactory('#' + chan,
        'YourMomDotCom', 2, chattiness=0.05))

I have found personally that, counter-intuitively, a length of 2 seems to produce the most realistic-looking responses. We've really accomplished a lot in a very little amount of code. Instead of my normal closing remarks, let me simply close with a snippet of conversation recently had with YourMomDotCom:

[9:57pm] ericflo: YourMomDotCom: Hello, I'd like to show you off for the readers of my blog.
[9:57pm] YourMomDotCom: ericflo: Hello, I'd like to do, after checking with another.
[9:57pm] ericflo: YourMomDotCom: I'm not exactly sure what you mean by that.
[9:57pm] YourMomDotCom: ericflo: I'm not sure, but I'm in "Easter mode" already I think.
[9:58pm] ericflo: YourMomDotCom: That's funny since we aren't really very close to easter.
[9:58pm] YourMomDotCom: ericflo: That's funny since i missed much of a ForeignKey field?
[9:58pm] ericflo: YourMomDotCom: I miss ForeignKey fields all the time
[9:58pm] YourMomDotCom: ericflo: I miss that
[9:58pm] ericflo: YourMomDotCom: We both agree, then.
[9:58pm] YourMomDotCom: ericflo: We both agree, then.
[9:58pm] ericflo: YourMomDotCom: yay!