diff --git a/README.md b/README.md
index 3b8dec5..52e54a6 100644
--- a/README.md
+++ b/README.md
@@ -47,6 +47,11 @@ with open("data.im", "wb") as f:
 # You can skip computing or checking CRCs, e.g. if your
 # embedded object already contains CRCs
 mb = MapBuffer(..., check_crc=False, compute_crc=False)
+
+# If your access pattern is such that the index and the 
+# download are similar in size (e.g. watershed meshes)
+# you can cache the index.
+mb = MapBuffer(..., index_cache="/tmp/helloworld.mbi")
 ```
 
 ## Installation
diff --git a/automated_test.py b/automated_test.py
index 2fac50d..d9d7b6c 100644
--- a/automated_test.py
+++ b/automated_test.py
@@ -3,11 +3,14 @@
 import mmap
 import os
 import random
+from unittest.mock import patch
 
 import numpy as np
 
 from mapbuffer import ValidationError, IntMap, MapBuffer, HEADER_LENGTH
 
+CACHE_PATH = "./test_index_cache.mbi"
+
 @pytest.mark.parametrize("compress", (None, "gzip", "br", "zstd", "lzma"))
 def test_empty(compress):
   mbuf = MapBuffer({}, compress=compress)
@@ -248,6 +251,96 @@ def test_set_object_intmap():
   except KeyError:
     pass
 
+@pytest.fixture(autouse=True)
+def cleanup_cache():
+    """Ensure cache file is removed before and after each test."""
+    if os.path.exists(CACHE_PATH):
+        os.remove(CACHE_PATH)
+    yield
+    if os.path.exists(CACHE_PATH):
+        os.remove(CACHE_PATH)
+
+
+def make_mapbuffer(data=None, **kwargs):
+    data = data or {1: b"hello", 2: b"world"}
+    return MapBuffer(data, index_cache=CACHE_PATH, **kwargs)
+
+
+def test_index_cache_file_is_created():
+    """Cache file should be written after first access."""
+    mbuf = make_mapbuffer()
+    mbuf.index()
+    assert os.path.exists(CACHE_PATH)
+
+
+def test_index_cache_header_and_index_written():
+    """Cache file should contain header + full index bytes."""
+    mbuf = make_mapbuffer()
+    index = mbuf.index()
+    
+    with open(CACHE_PATH, "rb") as f:
+        cached = f.read()
+
+    assert len(cached) == HEADER_LENGTH + index.nbytes
+
+
+def test_index_cache_is_loaded_from_disk():
+    """Second MapBuffer with same cache should read index from disk, not buffer."""
+    mbuf = make_mapbuffer()
+    original_index = mbuf.index().copy()
+
+    # Reload — this time the cache exists, so index should come from disk
+    mbuf2 = make_mapbuffer()
+    mbuf2._index = None  # ensure not inherited
+
+    with patch.object(np, "frombuffer", wraps=np.frombuffer) as mock_frombuffer:
+        loaded_index = mbuf2.index()
+        # np.frombuffer should NOT be called on the main buffer for the index
+        for call in mock_frombuffer.call_args_list:
+            args, kwargs = call
+            # Ensure we're not reading index from the primary buffer
+            assert kwargs.get("offset") != HEADER_LENGTH, \
+                "Index was re-read from buffer instead of cache"
+
+    np.testing.assert_array_equal(loaded_index, original_index)
+
+
+def test_index_cache_values_correct():
+    """Values retrieved using cache should match those from a non-cached buffer."""
+    mbuf_cached = make_mapbuffer()
+    mbuf_plain = MapBuffer({1: b"hello", 2: b"world"})
+
+    for key in [1, 2]:
+        assert mbuf_cached[key] == mbuf_plain[key]
+
+
+def test_crc_error_raised_despite_cache():
+    """CRC validation should still catch corruption even when cache exists."""
+    data = {1: b"hello", 2: b"world"}
+    mbuf = make_mapbuffer(data)
+    mbuf.index()  # populate cache
+
+    # Corrupt the data region in the buffer
+    buf = bytearray(mbuf.buffer)
+    idx = bytes(buf).index(b"hello")
+    buf[idx] = ord(b"H")
+    mbuf.buffer = bytes(buf)
+    mbuf._index = None  # force re-read so cache is used but data is still corrupt
+
+    with pytest.raises(ValidationError):
+        mbuf[1]
+
+
+def test_index_cache_not_rewritten_if_already_complete():
+    """Cache file should not be overwritten on second load."""
+    mbuf = make_mapbuffer()
+    mbuf.index()
 
+    mtime_after_first = os.path.getmtime(CACHE_PATH)
 
+    mbuf2 = make_mapbuffer()
+    mbuf2.index()
 
+    mtime_after_second = os.path.getmtime(CACHE_PATH)
+    assert mtime_after_first == mtime_after_second, \
+        "Cache file was unexpectedly rewritten on second access"
\ No newline at end of file
diff --git a/mapbuffer/mapbuffer.py b/mapbuffer/mapbuffer.py
index e7e7f5f..46fcad7 100644
--- a/mapbuffer/mapbuffer.py
+++ b/mapbuffer/mapbuffer.py
@@ -1,5 +1,6 @@
 from typing import Optional, Any, Union, Literal
 from collections.abc import Callable
+import os
 
 import mmap 
 import io
@@ -9,6 +10,7 @@
 from . import compression
 
 import crc32c
+import fasteners
 import numpy as np
 
 import mapbufferaccel
@@ -21,8 +23,8 @@ class MapBuffer:
   """Represents a usable int->bytes dictionary as a byte string."""
   __slots__ = (
     "data", "tobytesfn", "frombytesfn", 
-    "dtype", "buffer", "check_crc", "compute_crc",
-    "_header", "_index", "_compress"
+    "dtype", "buffer", "check_crc", "compute_crc", "index_cache",
+    "_header", "_index", "_compress", "_lock"
   )
   def __init__(
     self,
@@ -32,6 +34,7 @@ def __init__(
     frombytesfn:Optional[Callable[[bytes], Any]] = None,
     check_crc:bool = True, 
     compute_crc:bool = True,
+    index_cache:Optional[str] = None,
   ):
     """
     data: dict (int->byte serializable object) or bytes 
@@ -52,10 +55,14 @@ def __init__(
     self.buffer = None
     self.check_crc = check_crc
     self.compute_crc = compute_crc
+    self.index_cache = index_cache
 
     self._header = None
     self._index = None
     self._compress = None
+    self._lock = None
+    if self.index_cache is not None:
+      self._lock = fasteners.InterProcessReaderWriterLock(self.index_cache)
 
     if isinstance(data, dict):
       self.buffer = self.dict2buf(data, compress)
@@ -102,9 +109,29 @@ def header(self):
     if self._header is not None:
       return self._header
 
+    if self.index_cache is not None:
+      if os.path.exists(self.index_cache):
+        with self._lock.read_lock():
+          with open(self.index_cache, "rb") as f:
+            self._header = f.read(HEADER_LENGTH)
+
+        if len(self._header) == HEADER_LENGTH:
+          return self._header
+
     # seems dumb, buf if self.buffer is an object that
     # requires network access, this is a valuable cache
     self._header = self.buffer[:HEADER_LENGTH]
+
+    if self.index_cache is not None:
+      with self._lock.write_lock():
+        try:
+          if os.path.getsize(self.index_cache) < HEADER_LENGTH:
+            with open(self.index_cache, "wb") as f:
+              f.write(self._header)
+        except FileNotFoundError:
+          with open(self.index_cache, "wb") as f:
+            f.write(self._header)
+
     return self._header
 
   def index(self):
@@ -115,6 +142,18 @@ def index(self):
     N = len(self)
     index_length = 2 * N
 
+    if self.index_cache is not None:
+      try:
+        if os.path.getsize(self.index_cache) > HEADER_LENGTH:
+          with self._lock.read_lock():
+            with open(self.index_cache, "rb") as f:
+              f.seek(HEADER_LENGTH)
+              index = f.read(index_length * 8)
+          self._index = np.frombuffer(index, dtype=np.uint64).reshape((N,2))
+          return self._index
+      except FileNotFoundError:
+        pass
+
     if isinstance(self.buffer, (bytes,bytearray,np.ndarray,mmap.mmap)):
       self._index = np.frombuffer(
         self.buffer,
@@ -127,6 +166,15 @@ def index(self):
       index = self.buffer[HEADER_LENGTH:index_length+HEADER_LENGTH]
       self._index = np.frombuffer(index, dtype=np.uint64).reshape((N,2))
     
+    if self.index_cache is not None:
+      try:
+        if os.path.getsize(self.index_cache) == HEADER_LENGTH:
+          with self._lock.write_lock():
+            with open(self.index_cache, "ab") as f:
+              f.write(self._index.tobytes('C'))
+      except FileNotFoundError:
+        pass
+    
     return self._index
 
   def keys(self):
diff --git a/requirements.txt b/requirements.txt
index f3af510..8c9461d 100644
--- a/requirements.txt
+++ b/requirements.txt
@@ -1,6 +1,7 @@
 brotli
 crc32c
 deflate>=0.2.0
+fasteners
 numpy
 tqdm
 zstandard
\ No newline at end of file