diff --git a/CHANGELOG.md b/CHANGELOG.md
index 06b14b9..96ed0e1 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -1,5 +1,11 @@
 # Changelog
 
+## [Unreleased]
+
+### Changed
+
+- Remove the ``frozendict`` dependency, replacing it with a stdlib-only hashable ``dict`` subclass used to build ``lru_cache`` keys. This drops the last LGPL-3.0 dependency from einx without changing caching behaviour (https://github.com/fferflo/einx/pull/35).
+
 ## [0.4.3]
 
 ### Fixed
diff --git a/einx/_src/util/lru_cache.py b/einx/_src/util/lru_cache.py
index 6a168d2..c38d7b0 100644
--- a/einx/_src/util/lru_cache.py
+++ b/einx/_src/util/lru_cache.py
@@ -5,7 +5,6 @@
 from collections import defaultdict
 import numpy as np
 from functools import partial
-import frozendict
 import types
 
 _thread_local = threading.local()
@@ -14,13 +13,22 @@
 max_cache_size = int(os.environ.get("EINX_CACHE_SIZE", -1))
 
 
+class _FrozenDict(dict):
+    # A hashable dict used to build lru_cache keys. It still behaves like a
+    # mapping, since the frozen value is read back via key access downstream,
+    # but is hashable. Hashing is order-independent so that two equal dicts
+    # always produce the same cache key.
+    def __hash__(self):
+        return hash(frozenset(self.items()))
+
+
 def _freeze_value(x):
     if isinstance(x, np.ndarray):
         return _freeze_value(x.tolist())
     elif isinstance(x, list | tuple):
         return tuple(_freeze_value(x) for x in x)
     elif isinstance(x, dict):
-        return frozendict.frozendict({k: _freeze_value(v) for k, v in x.items()})
+        return _FrozenDict((k, _freeze_value(v)) for k, v in x.items())
     elif isinstance(x, types.SimpleNamespace):
         return _freeze_value(vars(x))
     elif isinstance(x, inspect.Parameter):
diff --git a/pyproject.toml b/pyproject.toml
index 6a98206..523321d 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -22,7 +22,6 @@ readme = "README.md"
 dependencies = [
     "numpy",
     "sympy",
-    "frozendict",
 ]
 
 [project.optional-dependencies]