diff --git a/klusta_pipeline/openephys_to_kwd.py b/klusta_pipeline/openephys_to_kwd.py
new file mode 100644
index 0000000..93bb2eb
--- /dev/null
+++ b/klusta_pipeline/openephys_to_kwd.py
@@ -0,0 +1,217 @@
+# Converts open ephys format files to kwd files
+import os
+import glob
+import numpy as np
+import h5py as h5
+
+# OpenEphys constants
+NUM_HEADER_BYTES = 1024
+SAMPLES_PER_RECORD = 1024
+BYTES_PER_SAMPLE = 2
+RECORD_SIZE = 4 + 8 + SAMPLES_PER_RECORD*BYTES_PER_SAMPLE + 10
+RECORD_RECORDING_OFFSET = 8+2
+RECORD_TIMESTAMP_OFFSET = 0
+
+
+def find_continuous_files(data_dir):
+    return glob.glob(os.path.join(data_dir, '*.continuous'))
+
+def find_events_files(data_dir):
+    return glob.glob(os.path.join(data_dir, '*.events'))
+
+def get_n_channels(continuous_files):
+    return len(continuous_files)
+
+def initialize_kwd_file(data_dir, experiment_name):
+    # Create an empty KWD file with the appropriate structure
+    kwd_filename = experiment_name + '.raw.kwd'
+    kwd_file = h5.File(os.path.join(data_dir, kwd_filename), "w-")
+    kwd_file.create_group("recordings")
+
+    continuous_files = find_continuous_files(data_dir)
+    unique_recording_numbers = get_openephys_unique_recording_numbers(continuous_files)
+    for rec_num in unique_recording_numbers:
+        kwd_file.create_group("/recordings/{}".format(rec_num))
+    return kwd_file
+
+def read_openephys_header(oe_file):
+    # return a dictionary of open ephys header values
+    oe_header = {}
+    oe_file.seek(0)
+    h = oe_file.read(1024).decode().replace('\n', '').replace('header.','')
+    for ind, header_item in enumerate(h.split(';')):
+        if '=' in header_item:
+            header_key = header_item.split(' = ')[0]
+            header_value = header_item.split(' = ')[1]
+            oe_header[header_key] = header_value
+    return oe_header
+
+def calculate_openephys_continuous_sizes(oe_file):
+    # Calculate: Number of Records, Number of Samples
+    n_file_bytes = os.fstat(oe_file.fileno()).st_size
+    n_record_bytes = n_file_bytes - NUM_HEADER_BYTES
+    
+    # Check if file is consistent
+    if n_record_bytes % RECORD_SIZE != 0:
+        raise Exception("File size inconsistent: possible corrupt data file")
+
+    n_records = n_record_bytes // RECORD_SIZE
+    n_samples = n_records * SAMPLES_PER_RECORD
+    return (n_records, n_samples)
+
+def load_openephys_continuous_record(oe_file, record_num):
+    # Loads a single record from an openephys continous file
+    
+    # Calculate start of record in bytes
+    record_start = 1024 + record_num*RECORD_SIZE
+    # Move to start of record
+    oe_file.seek(record_start)
+    # Get Timestamp
+    record_timestamp = np.fromfile(oe_file, np.dtype('<i8'), 1)
+    # Get Number of samples
+    record_n_samples = np.fromfile(oe_file, np.dtype('<u2'), 1)[0]
+    if record_n_samples != SAMPLES_PER_RECORD:
+        raise Exception('Found corrupted record in ' + str(record_num))
+    # Get recording number
+    record_recording_number = np.fromfile(oe_file, np.dtype('>u2'), 1)[0]
+    # Get raw record data
+    record_raw_data = np.fromfile(oe_file, np.dtype('>i2'), record_n_samples)
+    # Ignore record marker
+    oe_file.read(10)  
+    # Make array of timesamples
+    sample_numbers = np.arange(record_timestamp, 
+                               record_timestamp + record_n_samples)
+    # Make data array
+    data = np.zeros((record_n_samples, 2))
+    data[:, 0] = sample_numbers
+    data[:, 1] = record_raw_data
+    recording_start_sample = int(data[0, 0])
+
+    return (record_num, record_recording_number, recording_start_sample, data)
+
+def get_openephys_continuous_recording_numbers(oe_file):
+    # Returns an array of recording numbers within a .continuous file
+    recordings = []
+    (n_records, n_samples) = calculate_openephys_continuous_sizes(oe_file)
+    for record in range(n_records):
+        record_offset = 1024 + record*RECORD_SIZE
+        recording_offset = record_offset + RECORD_RECORDING_OFFSET
+        oe_file.seek(recording_offset)
+        record_recording_number = np.fromfile(oe_file, np.dtype('>u2'), 1)[0]
+        recordings.append(record_recording_number)
+    return recordings
+
+def get_openephys_continuous_timestamps(oe_file):
+    # Returns an array of recording numbers within a .continuous file
+    timestamps = []
+    (n_records, n_samples) = calculate_openephys_continuous_sizes(oe_file)
+    for record in range(n_records):
+        record_offset = 1024 + record*RECORD_SIZE
+        timestamp_offset = record_offset + RECORD_TIMESTAMP_OFFSET
+        oe_file.seek(timestamp_offset)
+        record_timestamp = np.fromfile(oe_file, np.dtype('<i8'), 1)
+        timestamps.append(record_timestamp)
+    return timestamps
+
+def get_openephys_continuous_record_metadata(oe_file):
+    # Returns an array of recording numbers within a .continuous file
+    recordings = []
+    timestamps = []
+    (n_records, n_samples) = calculate_openephys_continuous_sizes(oe_file)
+    for record in range(n_records):
+        record_offset = 1024 + record*RECORD_SIZE
+        timestamp_offset = record_offset + RECORD_TIMESTAMP_OFFSET
+        recording_offset = record_offset + RECORD_RECORDING_OFFSET
+        oe_file.seek(timestamp_offset)
+        record_timestamp = np.fromfile(oe_file, np.dtype('<i8'), 1)
+        timestamps.append(record_timestamp)
+        oe_file.seek(recording_offset)
+        record_recording_number = np.fromfile(oe_file, np.dtype('>u2'), 1)[0]
+        recordings.append(record_recording_number)
+    return (recordings, timestamps)
+
+def get_openephys_continuous_record_recording_number(oe_file, record):
+    offset = 1024 + record*RECORD_SIZE + RECORD_RECORDING_OFFSET
+    oe_file.seek(offset)
+    return np.fromfile(oe_file, np.dtype('>u2'), 1)[0]
+
+def get_openephys_continuous_record_timestamp(oe_file, record):
+    offset = 1024 + record*RECORD_SIZE + RECORD_TIMESTAMP_OFFSET
+    oe_file.seek(offset)
+    return np.fromfile(oe_file, np.dtype('<i8'), 1)
+
+def get_openephys_continuous_recording_timestamps(oe_file, recording):
+    (recordings, timestamps) = get_openephys_continuous_record_metadata(oe_file)
+    return timestamps[recordings == recording]
+
+def get_openephys_unique_recording_numbers(continuous_files):
+    recordings = []
+    for cf in continuous_files:
+        with open(cf, 'rb') as f:
+            cf_unique_recordings = get_openephys_continuous_recording_numbers(f)
+            recordings = recordings + cf_unique_recordings
+    return list(set(recordings))
+            
+def check_n_samples_consistency(continuous_files):
+    n_samples_list = []
+    for cf in continuous_files:
+        with open(cf, "rb") as oe_file:
+            n_samples = calculate_openephys_continuous_sizes(oe_file)[1]
+            n_samples_list.append(n_samples)
+    if len(set(n_samples_list)) != 1:
+        raise Exception("Inconsistent number of samples across channels")
+    return list(set(n_samples_list))[0]
+
+def get_openephys_channel_number(continuous_file):
+    cf_filename = os.path.split(continuous_file)[1]
+    cf_filename = cf_filename.replace('.continuous', '')
+    if 'CH' in cf_filename:
+        channel_number = int(cf_filename.split('CH')[-1]) - 1
+    if 'ADC' in cf_filename:
+        channel_number = int(cf_filename.split('ADC')[-1]) - 12
+    if 'AUX' in cf_filename:
+        channel_number = int(cf_filename.split('AUX')[-1]) - 4
+    return channel_number
+    
+def build_kwd_recording(kwd_file, continuous_files, recording_number):
+    n_channels = get_n_channels(continuous_files)
+    n_samples = check_n_samples_consistency(continuous_files)
+    dset = kwd_file.create_dataset("/recordings/{}/data".format(recording_number),
+                            (n_samples, n_channels), dtype='int16')
+
+    channel_bit_volts = np.zeros(n_channels)
+    channel_sample_rates = np.zeros(n_channels)
+    channel_timestamps = []
+
+    for cf in continuous_files:
+        # get channel number
+        channel = get_openephys_channel_number(cf)
+        print("Storing channel {} of {}".format(channel, n_channels))
+        with open(cf, "rb") as oe_file:
+            cf_header = read_openephys_header(oe_file)
+            channel_sample_rates[channel] = int(cf_header['sampleRate'])
+            channel_bit_volts[channel] = float(cf_header['bitVolts'])
+
+            (n_records, n_oe_file_samples) = calculate_openephys_continuous_sizes(oe_file)
+            (recordings, timestamps) = get_openephys_continuous_record_metadata(oe_file)
+            recordings = np.array(recordings)
+            timestamps = np.array(timestamps)
+            channel_timestamps.append(timestamps)
+
+            # Select records corresponding to this recording
+            recording_records = np.arange(n_records)[recordings==recording_number]
+            
+            # For each record, save the samples in the kwd recording dataset
+            for record_ind, record in enumerate(recording_records):
+                record_data_list = load_openephys_continuous_record(oe_file, record)
+                record_data = record_data_list[3] 
+                record_sample_lo = record_ind*SAMPLES_PER_RECORD
+                record_sample_hi = record_ind*SAMPLES_PER_RECORD + SAMPLES_PER_RECORD
+                #print(record_ind, record, record_sample_lo, record_sample_hi, n_records, n_samples)
+                dset[record_sample_lo:record_sample_hi, channel] = record_data[:, 1]
+    
+    # Save recording metadata
+    kwd_file.create_dataset("/recordings/{}/application_data/channel_sample_rates".format(recording_number), data=channel_sample_rates)
+    kwd_file.create_dataset("/recordings/{}/application_data/channel_bit_volts".format(recording_number), data=channel_bit_volts)
+    kwd_file.create_dataset("/recordings/{}/application_data/timestamps".format(recording_number), data=np.array(channel_timestamps))
+