diff --git a/README.md b/README.md
index ee1a5e2..6f7aeb4 100644
--- a/README.md
+++ b/README.md
@@ -6,6 +6,34 @@ This repo contains several algorithms for multi-hop reasoning on knowledge graph
 - [x] [Query2box](https://arxiv.org/abs/2002.05969)
 - [x] [GQE](https://arxiv.org/abs/1806.01445)
 
+## Installation
+
+Install directly from GitHub with:
+
+```shell
+$ pip install git+https://github.com/snap-stanford/KGReasoning.git
+```
+
+Install in development mode with:
+
+```shell
+$ git clone https://github.com/snap-stanford/KGReasoning.git
+$ cd KGReasoning
+$ pip install -e .
+```
+
+## Command Line Interface
+
+The `kgreasoning` command line tool is automatically installed. It can
+be used like in the following:
+
+```bash
+$ CUDA_VISIBLE_DEVICES=0 kgreasoning --cuda --do_train --do_valid --do_test \
+    --data_path data/FB15k-237-betae -n 128 -b 512 -d 800 -g 24 \
+    -lr 0.0001 --max_steps 450001 --cpu_num 1 --geo vec --valid_steps 15000 \
+    --tasks "1p.2p.3p.2i.3i.ip.pi.2u.up"
+```
+
 **KG Data**
 
 The KG data (FB15k, FB15k-237, NELL995) mentioned in the BetaE paper and the Query2box paper can be downloaded [here](http://snap.stanford.edu/betae/KG_data.zip). Note the two use the same training queries, but the difference is that the valid/test queries in BetaE paper have a maximum number of answers, making it more realistic.
diff --git a/example.sh b/example.sh
index 0f5b816..7cbe6a7 100755
--- a/example.sh
+++ b/example.sh
@@ -1,50 +1,50 @@
 
 # FB15k-237
-CUDA_VISIBLE_DEVICES=0 python main.py --cuda --do_train --do_valid --do_test \
+CUDA_VISIBLE_DEVICES=0 kgreasoning --cuda --do_train --do_valid --do_test \
   --data_path data/FB15k-237-betae -n 128 -b 512 -d 800 -g 24 \
   -lr 0.0001 --max_steps 450001 --cpu_num 1 --geo vec --valid_steps 15000 \
   --tasks "1p.2p.3p.2i.3i.ip.pi.2u.up"
 
-CUDA_VISIBLE_DEVICES=0 python main.py --cuda --do_train --do_valid --do_test \
+CUDA_VISIBLE_DEVICES=0 kgreasoning --cuda --do_train --do_valid --do_test \
   --data_path data/FB15k-237-betae -n 128 -b 512 -d 400 -g 60 \
   -lr 0.0001 --max_steps 450001 --cpu_num 1 --geo beta --valid_steps 15000 \
   -betam "(1600,2)"
 
-CUDA_VISIBLE_DEVICES=0 python main.py --cuda --do_train --do_valid --do_test \
+CUDA_VISIBLE_DEVICES=0 kgreasoning --cuda --do_train --do_valid --do_test \
   --data_path data/FB15k-237-betae -n 128 -b 512 -d 400 -g 24 \
   -lr 0.0001 --max_steps 450001 --cpu_num 1 --geo box --valid_steps 15000 \
   -boxm "(none,0.02)" --tasks "1p.2p.3p.2i.3i.ip.pi.2u.up"
 
 
 # FB15k
-CUDA_VISIBLE_DEVICES=0 python main.py --cuda --do_train --do_valid --do_test \
+CUDA_VISIBLE_DEVICES=0 kgreasoning --cuda --do_train --do_valid --do_test \
   --data_path data/FB15k-betae -n 128 -b 512 -d 800 -g 24 \
   -lr 0.0001 --max_steps 450001 --cpu_num 1 --geo vec --valid_steps 15000 \
   --tasks "1p.2p.3p.2i.3i.ip.pi.2u.up"
 
-CUDA_VISIBLE_DEVICES=0 python main.py --cuda --do_train --do_valid --do_test \
+CUDA_VISIBLE_DEVICES=0 kgreasoning --cuda --do_train --do_valid --do_test \
   --data_path data/FB15k-betae -n 128 -b 512 -d 400 -g 60 \
   -lr 0.0001 --max_steps 450001 --cpu_num 1 --geo beta --valid_steps 15000 \
   -betam "(1600,2)"
 
-CUDA_VISIBLE_DEVICES=0 python main.py --cuda --do_train --do_valid --do_test \
+CUDA_VISIBLE_DEVICES=0 kgreasoning --cuda --do_train --do_valid --do_test \
   --data_path data/FB15k-betae -n 128 -b 512 -d 400 -g 24 \
   -lr 0.0001 --max_steps 450001 --cpu_num 1 --geo box --valid_steps 15000 \
   -boxm "(none,0.02)" --tasks "1p.2p.3p.2i.3i.ip.pi.2u.up"
 
 
 # NELL
-CUDA_VISIBLE_DEVICES=0 python main.py --cuda --do_train --do_valid --do_test \
+CUDA_VISIBLE_DEVICES=0 kgreasoning --cuda --do_train --do_valid --do_test \
   --data_path data/NELL-betae -n 128 -b 512 -d 800 -g 24 \
   -lr 0.0001 --max_steps 450001 --cpu_num 1 --geo vec --valid_steps 15000 \
   --tasks "1p.2p.3p.2i.3i.ip.pi.2u.up"
 
-CUDA_VISIBLE_DEVICES=0 python main.py --cuda --do_train --do_valid --do_test \
+CUDA_VISIBLE_DEVICES=0 kgreasoning --cuda --do_train --do_valid --do_test \
   --data_path data/NELL-betae -n 128 -b 512 -d 400 -g 60 \
   -lr 0.0001 --max_steps 450001 --cpu_num 1 --geo beta --valid_steps 15000 \
   -betam "(1600,2)"
 
-CUDA_VISIBLE_DEVICES=0 python main.py --cuda --do_train --do_valid --do_test \
+CUDA_VISIBLE_DEVICES=0 kgreasoning --cuda --do_train --do_valid --do_test \
   --data_path data/NELL-betae -n 128 -b 512 -d 400 -g 24 \
   -lr 0.0001 --max_steps 450001 --cpu_num 1 --geo box --valid_steps 15000 \
   -boxm "(none,0.02)" --tasks "1p.2p.3p.2i.3i.ip.pi.2u.up"
@@ -52,7 +52,7 @@ CUDA_VISIBLE_DEVICES=0 python main.py --cuda --do_train --do_valid --do_test \
 
 ## Evaluation
 
-CUDA_VISIBLE_DEVICES=0 python main.py --cuda --do_test \
+CUDA_VISIBLE_DEVICES=0 kgreasoning --cuda --do_test \
   --data_path data/FB15k-237-betae -n 128 -b 512 -d 400 -g 24 \
   -lr 0.0001 --max_steps 450001 --cpu_num 1 --geo box --valid_steps 15000 \
   -boxm "(none,0.02)" --tasks "1p.2p.3p.2i.3i.ip.pi.2u.up" --checkpoint_path $CKPT_PATH
diff --git a/setup.cfg b/setup.cfg
new file mode 100644
index 0000000..26c9bf6
--- /dev/null
+++ b/setup.cfg
@@ -0,0 +1,51 @@
+##########################
+# Setup.py Configuration #
+##########################
+[metadata]
+name = kgreasoning
+version = 0.0.1-dev
+
+# URLs associated with the project
+url = https://github.com/snap-stanford/KGReasoning
+download_url = https://github.com/snap-stanford/KGReasoning/releases
+project_urls =
+    Bug Tracker = https://github.com/snap-stanford/KGReasoning/issues
+    Source Code = https://github.com/snap-stanford/KGReasoning
+
+# Author information
+author = Hongyu Ren
+author_email = hyren@cs.stanford.edu
+
+# License Information
+license = MIT
+license_file = LICENSE
+
+# Search tags
+classifiers =
+    Development Status :: 1 - Planning
+    License :: OSI Approved :: MIT License
+    Operating System :: OS Independent
+    Programming Language :: Python
+
+[options]
+install_requires =
+    torch
+    tqdm
+    tensorboardX
+
+# Random options
+zip_safe = false
+include_package_data = True
+python_requires = >=3.6
+
+# Where is my code
+packages = find:
+package_dir =
+    = src
+
+[options.packages.find]
+where = src
+
+[options.entry_points]
+console_scripts =
+    kgreasoning = kgreasoning.main:main
diff --git a/setup.py b/setup.py
new file mode 100644
index 0000000..a78fbfd
--- /dev/null
+++ b/setup.py
@@ -0,0 +1,8 @@
+# -*- coding: utf-8 -*-
+
+"""Setup module."""
+
+import setuptools
+
+if __name__ == '__main__':
+    setuptools.setup()
diff --git a/src/kgreasoning/__init__.py b/src/kgreasoning/__init__.py
new file mode 100644
index 0000000..e69de29
diff --git a/dataloader.py b/src/kgreasoning/dataloader.py
similarity index 95%
rename from dataloader.py
rename to src/kgreasoning/dataloader.py
index 5f4f953..9c3230e 100755
--- a/dataloader.py
+++ b/src/kgreasoning/dataloader.py
@@ -8,7 +8,7 @@
 import torch
 
 from torch.utils.data import Dataset
-from util import list2tuple, tuple2list, flatten
+from kgreasoning.util import list2tuple, tuple2list, flatten
 
 class TestDataset(Dataset):
     def __init__(self, queries, nentity, nrelation):
diff --git a/main.py b/src/kgreasoning/main.py
similarity index 96%
rename from main.py
rename to src/kgreasoning/main.py
index 6b76007..ad83cd0 100755
--- a/main.py
+++ b/src/kgreasoning/main.py
@@ -13,14 +13,14 @@
 import numpy as np
 import torch
 from torch.utils.data import DataLoader
-from models import KGReasoning
-from dataloader import TestDataset, TrainDataset, SingledirectionalOneShotIterator
+from kgreasoning.models import KGReasoning
+from kgreasoning.dataloader import TestDataset, TrainDataset, SingledirectionalOneShotIterator
 from tensorboardX import SummaryWriter
 import time
 import pickle
 from collections import defaultdict
 from tqdm import tqdm
-from util import flatten_query, list2tuple, parse_time, set_global_seed, eval_tuple
+from kgreasoning.util import flatten_query, list2tuple, parse_time, set_global_seed, eval_tuple
 
 query_name_dict = {('e',('r',)): '1p', 
                     ('e', ('r', 'r')): '2p',
@@ -193,7 +193,7 @@ def load_data(args, tasks):
 
     return train_queries, train_answers, valid_queries, valid_hard_answers, valid_easy_answers, test_queries, test_hard_answers, test_easy_answers
 
-def main(args):
+def main_helper(args):
     set_global_seed(args.seed)
     tasks = args.tasks.split('.')
     for task in tasks:
@@ -445,5 +445,10 @@ def main(args):
 
     logging.info("Training finished!!")
 
+
+def main():
+    main_helper(parse_args())
+
+
 if __name__ == '__main__':
-    main(parse_args())
\ No newline at end of file
+    main()
diff --git a/models.py b/src/kgreasoning/models.py
similarity index 97%
rename from models.py
rename to src/kgreasoning/models.py
index b761009..524b3b3 100755
--- a/models.py
+++ b/src/kgreasoning/models.py
@@ -10,7 +10,7 @@
 import torch.nn as nn
 import torch.nn.functional as F
 from torch.utils.data import DataLoader
-from dataloader import TestDataset, TrainDataset, SingledirectionalOneShotIterator
+from kgreasoning.dataloader import TestDataset, TrainDataset, SingledirectionalOneShotIterator
 import random
 import pickle
 import math
diff --git a/util.py b/src/kgreasoning/util.py
similarity index 100%
rename from util.py
rename to src/kgreasoning/util.py