feat(tools): support callable class instances and bound methods

src/anthropic/lib/tools/_beta_functions.py

@@ -3,7 +3,7 @@
 import logging
 from abc import ABC, abstractmethod
 from typing import Any, Union, Generic, TypeVar, Callable, Iterable, Coroutine, cast, overload
-from inspect import iscoroutinefunction
+from inspect import ismethod, isfunction, iscoroutinefunction
 from typing_extensions import TypeAlias, override
 
 import pydantic
@@ -23,6 +23,33 @@
 
 BetaFunctionToolResultType: TypeAlias = Union[str, Iterable[BetaContent]]
 
+
+def _normalize_callable(func: Callable[..., Any]) -> Callable[..., Any]:
+    """Normalize a callable to a function that can be used with pydantic.validate_call.
+
+    If the callable is a class instance with a __call__ method (but not a function or method),
+    this extracts the bound __call__ method. This allows callable class instances to be used
+    as tools without requiring manual extraction of __call__.
+
+    Args:
+        func: A function, method, or callable instance
+
+    Returns:
+        A function or bound method suitable for use with pydantic.validate_call
+    """
+    # If it's already a function or method, use it directly
+    if isfunction(func) or ismethod(func):
+        return func
+
+    # If it's a callable instance (class with __call__), extract the bound __call__ method
+    if callable(func):
+        call_method = func.__call__  # pyright: ignore[reportFunctionMemberAccess]  # noqa: B004
+        if ismethod(call_method):
+            return call_method
+
+    return func
+
+
 Function = Callable[..., BetaFunctionToolResultType]
 FunctionT = TypeVar("FunctionT", bound=Function)
 
@@ -83,9 +110,11 @@ def __init__(
         if _compat.PYDANTIC_V1:
             raise RuntimeError("Tool functions are only supported with Pydantic v2")
 
-        self.func = func
-        self._func_with_validate = pydantic.validate_call(func)
-        self.name = name or func.__name__
+        # Normalize callable instances to their __call__ method
+        normalized_func = _normalize_callable(func)
+        self.func = cast(CallableT, normalized_func)
+        self._func_with_validate = pydantic.validate_call(normalized_func)
+        self.name = name or normalized_func.__name__
         self._defer_loading = defer_loading
 
         self.description = description or self._get_description_from_docstring()

tests/lib/tools/test_functions.py

@@ -439,6 +439,80 @@ def simple_add(a: int, b: int) -> str:
         assert function_tool.input_schema == expected_schema
 
 
+    def test_callable_class_instance(self) -> None:
+        """Test that callable class instances can be used as tools."""
+
+        class FetchProduct:
+            def __init__(self, ctx: dict[str, str]) -> None:
+                self.ctx = ctx
+
+            def __call__(self, product_id: int) -> str:
+                """Fetch a product by ID."""
+                return f"Product {product_id} from {self.ctx['session']}"
+
+        instance = FetchProduct({"session": "test-session"})
+        tool = beta_tool(instance, name="fetch_product")
+
+        assert tool.name == "fetch_product"
+        assert tool.description == "Fetch a product by ID."
+        assert tool.call({"product_id": 123}) == "Product 123 from test-session"
+
+        # Check schema
+        expected_schema = {
+            "additionalProperties": False,
+            "type": "object",
+            "properties": {
+                "product_id": {"title": "Product Id", "type": "integer"},
+            },
+            "required": ["product_id"],
+        }
+        assert tool.input_schema == expected_schema
+
+    def test_bound_method(self) -> None:
+        """Test that bound methods can be used as tools."""
+
+        class WeatherService:
+            def __init__(self, api_key: str) -> None:
+                self.api_key = api_key
+
+            def get_weather(self, location: str) -> str:
+                """Get weather for a location."""
+                return f"Weather in {location} (using key: {self.api_key[:4]}...)"
+
+        service = WeatherService("secret-api-key")
+        tool = beta_tool(service.get_weather, name="weather")
+
+        assert tool.name == "weather"
+        assert tool.description == "Get weather for a location."
+        assert tool.call({"location": "London"}) == "Weather in London (using key: secr...)"
+
+        # Check schema
+        expected_schema = {
+            "additionalProperties": False,
+            "type": "object",
+            "properties": {
+                "location": {"title": "Location", "type": "string"},
+            },
+            "required": ["location"],
+        }
+        assert tool.input_schema == expected_schema
+
+    def test_callable_class_without_explicit_name(self) -> None:
+        """Test that callable class instances infer name from __call__ method."""
+
+        class MyTool:
+            def __call__(self, x: int) -> str:
+                """Process x."""
+                return str(x)
+
+        instance = MyTool()
+        tool = beta_tool(instance)
+
+        # Should use __call__ as the name since that's the actual method
+        assert tool.name == "__call__"
+        assert tool.description == "Process x."
+
+
 def _get_parameters_info(fn: BaseFunctionTool[Any]) -> dict[str, str]:
     param_info: dict[str, str] = {}
     for param in fn._parsed_docstring.params:

tests/lib/tools/test_runners.py

@@ -547,6 +547,86 @@ def tool_runner(client: Anthropic) -> BetaToolRunner[None]:
             respx_mock=respx_mock,
         )
 
+    @pytest.mark.respx(base_url=base_url)
+    def test_callable_class_instance_tool(
+        self, client: Anthropic, respx_mock: MockRouter, monkeypatch: pytest.MonkeyPatch
+    ) -> None:
+        """Test that callable class instances work with tool_runner."""
+
+        class WeatherTool:
+            def __init__(self, default_units: str) -> None:
+                self.default_units = default_units
+
+            def __call__(self, location: str, units: Literal["c", "f"] = "c") -> str:
+                """Lookup the weather for a given city in either celsius or fahrenheit
+
+                Args:
+                    location: The city and state, e.g. San Francisco, CA
+                    units: Unit for the output, either 'c' for celsius or 'f' for fahrenheit
+                Returns:
+                    A dictionary containing the location, temperature, and weather condition.
+                """
+                actual_units = units or self.default_units
+                return json.dumps(_get_weather(location, actual_units))
+
+        weather_instance = WeatherTool(default_units="f")
+        weather_tool = beta_tool(weather_instance, name="get_weather")
+
+        message = make_snapshot_request(
+            lambda c: c.beta.messages.tool_runner(
+                max_tokens=1024,
+                model="claude-haiku-4-5",
+                tools=[weather_tool],
+                messages=[{"role": "user", "content": "What is the weather in SF?"}],
+            ).until_done(),
+            content_snapshot=snapshots["basic"]["responses"],
+            path="/v1/messages",
+            mock_client=client,
+            respx_mock=respx_mock,
+        )
+
+        assert print_obj(message, monkeypatch) == snapshots["basic"]["result"]
+
+    @pytest.mark.respx(base_url=base_url)
+    def test_bound_method_tool(
+        self, client: Anthropic, respx_mock: MockRouter, monkeypatch: pytest.MonkeyPatch
+    ) -> None:
+        """Test that bound methods work with tool_runner."""
+
+        class WeatherService:
+            def __init__(self, api_key: str) -> None:
+                self.api_key = api_key
+
+            def get_weather(self, location: str, units: Literal["c", "f"]) -> str:
+                """Lookup the weather for a given city in either celsius or fahrenheit
+
+                Args:
+                    location: The city and state, e.g. San Francisco, CA
+                    units: Unit for the output, either 'c' for celsius or 'f' for fahrenheit
+                Returns:
+                    A dictionary containing the location, temperature, and weather condition.
+                """
+                # In a real scenario, self.api_key would be used
+                return json.dumps(_get_weather(location, units))
+
+        service = WeatherService(api_key="secret-key")
+        weather_tool = beta_tool(service.get_weather, name="get_weather")
+
+        message = make_snapshot_request(
+            lambda c: c.beta.messages.tool_runner(
+                max_tokens=1024,
+                model="claude-haiku-4-5",
+                tools=[weather_tool],
+                messages=[{"role": "user", "content": "What is the weather in SF?"}],
+            ).until_done(),
+            content_snapshot=snapshots["basic"]["responses"],
+            path="/v1/messages",
+            mock_client=client,
+            respx_mock=respx_mock,
+        )
+
+        assert print_obj(message, monkeypatch) == snapshots["basic"]["result"]
+
 
 @pytest.mark.skipif(PYDANTIC_V1, reason="tool runner not supported with pydantic v1")
 @pytest.mark.respx(base_url=base_url)