lm-deluge 0.0.67__py3-none-any.whl → 0.0.90__py3-none-any.whl

lm_deluge/tool/__init__.py

@@ -0,0 +1,1130 @@
+import asyncio
+import inspect
+from concurrent.futures import ThreadPoolExecutor
+from functools import lru_cache
+from typing import (
+    Annotated,
+    Any,
+    Callable,
+    Coroutine,
+    Literal,
+    TypedDict,
+    get_args,
+    get_origin,
+    get_type_hints,
+)
+
+from fastmcp import Client  # pip install fastmcp >= 2.0
+from mcp.types import Tool as MCPTool
+from pydantic import BaseModel, Field, TypeAdapter, field_validator
+
+from lm_deluge.image import Image
+from lm_deluge.prompt import Text, ToolResultPart
+
+
+@lru_cache(maxsize=1000)
+def _get_cached_typeadapter(cls: type | Callable) -> TypeAdapter:
+    """
+    Cache TypeAdapters since they're expensive to create.
+    For functions, we also handle Annotated[T, "string"] -> Annotated[T, Field(description="string")].
+    """
+    if inspect.isfunction(cls) or inspect.ismethod(cls):
+        if hasattr(cls, "__annotations__") and cls.__annotations__:
+            try:
+                resolved_hints = get_type_hints(cls, include_extras=True)
+            except Exception:
+                resolved_hints = cls.__annotations__
+
+            # Convert Annotated[T, "string"] to Annotated[T, Field(description="string")]
+            processed_hints = {}
+            for name, annotation in resolved_hints.items():
+                if (
+                    get_origin(annotation) is Annotated
+                    and len(get_args(annotation)) == 2
+                    and isinstance(get_args(annotation)[1], str)
+                ):
+                    base_type, description = get_args(annotation)
+                    processed_hints[name] = Annotated[
+                        base_type, Field(description=description)
+                    ]
+                else:
+                    processed_hints[name] = annotation
+
+            # Create new function with processed annotations if changed
+            if processed_hints != cls.__annotations__:
+                import types
+
+                if inspect.ismethod(cls):
+                    actual_func = cls.__func__
+                    code = actual_func.__code__
+                    globals_dict = actual_func.__globals__
+                    name = actual_func.__name__
+                    defaults = actual_func.__defaults__
+                    kwdefaults = actual_func.__kwdefaults__
+                    closure = actual_func.__closure__
+                else:
+                    code = cls.__code__
+                    globals_dict = cls.__globals__
+                    name = cls.__name__
+                    defaults = cls.__defaults__
+                    kwdefaults = cls.__kwdefaults__
+                    closure = cls.__closure__
+
+                new_func = types.FunctionType(
+                    code,
+                    globals_dict,
+                    name,
+                    defaults,
+                    closure,
+                )
+                if kwdefaults is not None:
+                    new_func.__kwdefaults__ = kwdefaults
+                new_func.__dict__.update(cls.__dict__)
+                new_func.__module__ = cls.__module__
+                new_func.__qualname__ = getattr(cls, "__qualname__", cls.__name__)
+                new_func.__annotations__ = processed_hints
+
+                if inspect.ismethod(cls):
+                    new_method = types.MethodType(new_func, cls.__self__)
+                    return TypeAdapter(new_method)
+                else:
+                    return TypeAdapter(new_func)
+
+    return TypeAdapter(cls)
+
+
+def _clean_schema(
+    schema: dict[str, Any],
+    *,
+    prune_titles: bool = True,
+    prune_additional_properties: bool = True,
+) -> dict[str, Any]:
+    """
+    Clean up a JSON schema by removing titles and additionalProperties: false.
+    This is applied recursively to all nested schemas.
+    """
+
+    def _traverse(node: Any) -> Any:
+        if isinstance(node, dict):
+            new_node = {}
+            for key, value in node.items():
+                # Skip titles if pruning
+                if prune_titles and key == "title":
+                    continue
+                # Skip additionalProperties: false if pruning
+                if (
+                    prune_additional_properties
+                    and key == "additionalProperties"
+                    and value is False
+                ):
+                    continue
+                new_node[key] = _traverse(value)
+            return new_node
+        elif isinstance(node, list):
+            return [_traverse(item) for item in node]
+        else:
+            return node
+
+    return _traverse(schema)
+
+
+def _get_type_hint_string(type_annotation: Any) -> str:
+    """
+    Get a readable string representation of a type annotation.
+    Handles generic types, unions, etc.
+    """
+    import re
+
+    # Handle None type
+    if type_annotation is type(None):
+        return "None"
+
+    # For generic types, get_origin and get_args give us the components
+    origin = get_origin(type_annotation)
+    args = get_args(type_annotation)
+
+    if origin is not None and args:
+        # Get the origin name (list, dict, etc.)
+        if hasattr(origin, "__name__"):
+            origin_name = origin.__name__
+        else:
+            origin_name = str(origin).replace("typing.", "")
+
+        # Recursively get arg strings
+        arg_strs = [_get_type_hint_string(arg) for arg in args]
+
+        # Handle Union types (including | syntax)
+        if origin_name in ("Union", "UnionType"):
+            return " | ".join(arg_strs)
+
+        return f"{origin_name}[{', '.join(arg_strs)}]"
+
+    # Try to get __name__ for simple types (int, str, custom classes)
+    if hasattr(type_annotation, "__name__"):
+        return type_annotation.__name__
+
+    # For anything else, use string representation and clean it up
+    type_str = str(type_annotation)
+
+    # Remove module prefixes like '__main__.', 'mymodule.', etc.
+    type_str = re.sub(r"\b\w+\.", "", type_str)
+    # Remove 'typing.' prefix (in case it's still there)
+    type_str = type_str.replace("typing.", "")
+    # Remove 'typing_extensions.' prefix
+    type_str = type_str.replace("typing_extensions.", "")
+
+    return type_str
+
+
+def _format_output_schema_for_description(
+    return_type: Any,
+    output_schema: dict[str, Any] | None,
+) -> str | None:
+    """
+    Format output schema information for inclusion in tool description.
+
+    Returns a string like:
+        "Returns: list[SearchResult]
+
+        SearchResult: {"properties": {...}, "type": "object"}"
+
+    Or None if there's no meaningful output schema to show.
+    """
+    import json
+
+    if return_type is None or return_type is inspect.Parameter.empty:
+        return None
+
+    # Get the type hint string
+    type_hint = _get_type_hint_string(return_type)
+
+    # Start with the return type
+    parts = [f"Returns: {type_hint}"]
+
+    # If there are $defs, include them
+    if output_schema and "$defs" in output_schema:
+        defs = output_schema["$defs"]
+        for def_name, def_schema in defs.items():
+            # Format the schema compactly (single line)
+            schema_str = json.dumps(def_schema, separators=(",", ":"))
+            parts.append(f"{def_name}: {schema_str}")
+
+    return "\n\n".join(parts)
+
+
+def _is_typeddict(cls: Any) -> bool:
+    """Check if a class is a TypedDict."""
+    return (
+        isinstance(cls, type)
+        and hasattr(cls, "__annotations__")
+        and hasattr(cls, "__total__")
+    )
+
+
+def _normalize_parameters(
+    params: Any,
+) -> tuple[dict[str, Any], list[str], dict[str, Any] | None]:
+    """
+    Normalize various parameter input formats to JSON schema components.
+
+    Accepts:
+        - None -> empty schema
+        - dict with "type" keys (already JSON schema) -> pass through
+        - dict mapping names to Python types {name: str, age: int}
+        - dict mapping names to (type, extras) tuples {name: (str, {"description": "..."})}
+        - Pydantic BaseModel class
+        - TypedDict class
+
+    Returns:
+        (properties, required, definitions)
+    """
+
+    def _schema_from_type(annotation: Any) -> dict[str, Any]:
+        """
+        Prefer TypeAdapter-based schemas (handles Union/Optional, Annotated, etc).
+        Fall back to the legacy mapper if TypeAdapter cannot handle the type.
+        """
+        try:
+            ta = TypeAdapter(annotation)
+            return _clean_schema(ta.json_schema())
+        except Exception:
+            return _python_type_to_json_schema(annotation)
+
+    if params is None:
+        return {}, [], None
+
+    # Pydantic model
+    if isinstance(params, type) and issubclass(params, BaseModel):
+        schema = params.model_json_schema()
+        schema = _clean_schema(schema)
+        properties = schema.get("properties", {})
+        required = schema.get("required", [])
+        definitions = schema.get("$defs")
+        return properties, required, definitions
+
+    # TypedDict
+    if _is_typeddict(params):
+        try:
+            ta = TypeAdapter(params)
+            schema = _clean_schema(ta.json_schema())
+            properties = schema.get("properties", {})
+            required = schema.get("required", [])
+            definitions = schema.get("$defs")
+            return properties, required, definitions
+        except Exception:
+            # Fall back to manual extraction
+            hints = get_type_hints(params)
+            properties = {}
+            required = []
+            for field_name, field_type in hints.items():
+                properties[field_name] = _python_type_to_json_schema(field_type)
+                required.append(field_name)
+            return properties, required, None
+
+    # Must be a dict at this point
+    if not isinstance(params, dict):
+        raise TypeError(
+            f"parameters must be a dict, Pydantic model, or TypedDict, "
+            f"got {type(params).__name__}"
+        )
+
+    # Check if it's already a JSON schema (has "type" keys in values)
+    # vs a simple {name: type} mapping
+    if params and all(
+        isinstance(v, dict) and "type" in v for v in params.values() if v is not None
+    ):
+        # Already JSON schema format - extract required from presence of "optional" key
+        required = [
+            name for name, schema in params.items() if not schema.get("optional", False)
+        ]
+        # Remove "optional" keys as they're not valid JSON schema
+        cleaned = {}
+        for name, schema in params.items():
+            cleaned[name] = {k: v for k, v in schema.items() if k != "optional"}
+        return cleaned, required, None
+
+    # Simple {name: type} or {name: (type, extras)} mapping
+    properties = {}
+    required = []
+
+    for param_name, param_spec in params.items():
+        # Tuple of (type, extras)
+        if isinstance(param_spec, tuple) and len(param_spec) == 2:
+            param_type, extras = param_spec
+            if isinstance(extras, dict):
+                schema = _schema_from_type(param_type)
+                schema.update(extras)
+                # Remove "optional" key as it's not valid JSON schema
+                is_optional = schema.pop("optional", False)
+                properties[param_name] = schema
+                if not is_optional:
+                    required.append(param_name)
+                continue
+
+        # Python type (int, str, list[str], etc.)
+        if isinstance(param_spec, type) or get_origin(param_spec) is not None:
+            properties[param_name] = _schema_from_type(param_spec)
+            required.append(param_name)
+            continue
+
+        # Already a JSON schema dict
+        if isinstance(param_spec, dict):
+            schema = param_spec.copy()
+            is_optional = schema.pop("optional", False)
+            properties[param_name] = schema
+            if not is_optional:
+                required.append(param_name)
+            continue
+
+        # Unknown - try to convert
+        properties[param_name] = _schema_from_type(param_spec)
+        required.append(param_name)
+
+    return properties, required, None
+
+
+def _python_type_to_json_schema(python_type: Any) -> dict[str, Any]:
+    """
+    Convert Python type annotations to JSON Schema.
+    Handles: primitives, Optional, Literal, list[T], dict[str, T], Union.
+    """
+    # Get origin and args for generic types
+    origin = get_origin(python_type)
+    args = get_args(python_type)
+
+    # Handle Optional[T] or T | None
+    if origin is type(None) or python_type is type(None):
+        return {"type": "null"}
+
+    # Handle Union types (including Optional)
+    if origin is Literal:
+        # Literal["a", "b"] -> enum
+        return {"type": "string", "enum": list(args)}
+
+    # Handle list[T]
+    if origin is list:
+        if args:
+            items_schema = _python_type_to_json_schema(args[0])
+            return {"type": "array", "items": items_schema}
+        return {"type": "array"}
+
+    # Handle dict[str, T]
+    if origin is dict:
+        if len(args) >= 2:
+            # For dict[str, T], we can set additionalProperties
+            value_schema = _python_type_to_json_schema(args[1])
+            return {"type": "object", "additionalProperties": value_schema}
+        return {"type": "object"}
+
+    # Handle basic types
+    if python_type is int:
+        return {"type": "integer"}
+    elif python_type is float:
+        return {"type": "number"}
+    elif python_type is str:
+        return {"type": "string"}
+    elif python_type is bool:
+        return {"type": "boolean"}
+    elif python_type is list:
+        return {"type": "array"}
+    elif python_type is dict:
+        return {"type": "object"}
+    else:
+        # Default to string for unknown types
+        return {"type": "string"}
+
+
+async def _load_all_mcp_tools(client: Client) -> list["Tool"]:
+    metas: list[MCPTool] = await client.list_tools()
+
+    def make_runner(name: str):
+        async def _async_call(**kw):
+            async with client:
+                # maybe should be call_tool_mcp if don't want to raise error
+                raw_result = await client.call_tool(name, kw)
+
+                # for now just concatenate them all into a result string
+                results = []
+                if not isinstance(raw_result, list):  # newer versions of fastmcp
+                    content_blocks = raw_result.content
+                else:
+                    content_blocks = raw_result
+                for block in content_blocks:
+                    if block.type == "text":
+                        results.append(Text(block.text))
+                    elif block.type == "image":
+                        data_url = f"data:{block.mimeType};base64,{block.data}"
+                        results.append(Image(data=data_url))
+
+                return results
+
+        return _async_call
+
+    tools: list[Tool] = []
+    for m in metas:
+        # Extract definitions from the schema (could be $defs or definitions)
+        definitions = m.inputSchema.get("$defs") or m.inputSchema.get("definitions")
+
+        tools.append(
+            Tool(
+                name=m.name,
+                description=m.description,
+                parameters=m.inputSchema.get("properties", {}),
+                required=m.inputSchema.get("required", []),
+                additionalProperties=m.inputSchema.get("additionalProperties"),
+                definitions=definitions,
+                run=make_runner(m.name),
+            )
+        )
+    return tools
+
+
+class Tool(BaseModel):
+    """
+    Provider‑agnostic tool definition with no extra nesting.
+
+    The `parameters` argument accepts multiple formats:
+        - dict with JSON schema: {"query": {"type": "string"}}
+        - dict with Python types: {"query": str, "limit": int}
+        - dict with (type, extras) tuples: {"query": (str, {"description": "..."})}
+        - Pydantic BaseModel class
+        - TypedDict class
+
+    Examples:
+        # From JSON schema (traditional)
+        Tool(name="search", parameters={"query": {"type": "string"}}, ...)
+
+        # From Python types (simple)
+        Tool(name="search", parameters={"query": str, "limit": int}, ...)
+
+        # From Pydantic model
+        class SearchParams(BaseModel):
+            query: str
+            limit: int = 10
+        Tool(name="search", parameters=SearchParams, ...)
+
+        # From TypedDict
+        class SearchParams(TypedDict):
+            query: str
+            limit: NotRequired[int]
+        Tool(name="search", parameters=SearchParams, ...)
+
+        # From function (recommended for most cases)
+        Tool.from_function(my_search_function)
+    """
+
+    model_config = {"arbitrary_types_allowed": True}
+
+    name: str
+    description: str | None = None
+    parameters: dict[str, Any] | None = None
+    required: list[str] = Field(default_factory=list)
+    additionalProperties: bool | None = None
+    # if desired, can provide a callable to run the tool
+    run: Callable | None = None
+    # for built-in tools that don't require schema
+    is_built_in: bool = False
+    type: str | None = None
+    built_in_args: dict[str, Any] = Field(default_factory=dict)
+    # JSON Schema definitions (for $ref support)
+    definitions: dict[str, Any] | None = None
+    # Output schema (extracted from return type annotation)
+    output_schema: dict[str, Any] | None = None
+    # TypeAdapter for output validation (not serialized, stored as private attr)
+    _output_type_adapter: TypeAdapter | None = None
+
+    def __init__(self, **data):
+        # Normalize parameters before passing to Pydantic
+        raw_params = data.get("parameters")
+        if raw_params is not None:
+            properties, required_fields, definitions = _normalize_parameters(raw_params)
+            data["parameters"] = properties
+            # Only set required if not explicitly provided (check for key presence, not truthiness)
+            if "required" not in data:
+                data["required"] = required_fields
+            # Only set definitions if not explicitly provided and we have new ones
+            if definitions and "definitions" not in data:
+                data["definitions"] = definitions
+
+        super().__init__(**data)
+
+    @field_validator("name")
+    @classmethod
+    def validate_name(cls, v: str) -> str:
+        if v.startswith("_computer_"):
+            raise ValueError(
+                f"Tool name '{v}' uses reserved prefix '_computer_'. "
+                "This prefix is reserved for computer use actions."
+            )
+        return v
+
+    def _is_async(self) -> bool:
+        return inspect.iscoroutinefunction(self.run)
+
+    def _validate_output(self, result: Any) -> Any:
+        """Validate output against output_schema if TypeAdapter is available."""
+        if self._output_type_adapter is None:
+            raise ValueError(
+                "Cannot validate output: no output type adapter available. "
+                "Make sure the tool was created with from_function() and has a return type annotation."
+            )
+        # This will raise ValidationError if result doesn't match the schema
+        return self._output_type_adapter.validate_python(result)
+
+    def call(
+        self, *, validate_output: bool = False, **kwargs
+    ) -> str | list[ToolResultPart]:
+        """
+        Call the tool with the given arguments.
+
+        Args:
+            validate_output: If True, validate the return value against the
+                output schema. Raises ValidationError if validation fails.
+                Requires the tool to have been created with from_function()
+                and have a return type annotation.
+            **kwargs: Arguments to pass to the tool function.
+
+        Returns:
+            The result of the tool function.
+
+        Raises:
+            ValueError: If no run function is provided or validation is requested
+                but no output type adapter is available.
+            pydantic.ValidationError: If validate_output=True and the result
+                doesn't match the output schema.
+        """
+        if self.run is None:
+            raise ValueError("No run function provided")
+
+        if self._is_async():
+            coro: Coroutine = self.run(**kwargs)  # type: ignore[arg-type]
+            try:
+                loop = asyncio.get_running_loop()
+                assert loop
+            except RuntimeError:
+                # no loop → safe to block
+                result = asyncio.run(coro)
+            else:
+                # Loop is running → execute coroutine in a worker thread
+                def _runner():
+                    return asyncio.run(coro)
+
+                with ThreadPoolExecutor(max_workers=1) as executor:
+                    result = executor.submit(_runner).result()
+        else:
+            # plain function
+            result = self.run(**kwargs)
+
+        if validate_output:
+            self._validate_output(result)
+
+        return result
+
+    async def acall(
+        self, *, validate_output: bool = False, **kwargs
+    ) -> str | list[ToolResultPart]:
+        """
+        Async version of call().
+
+        Args:
+            validate_output: If True, validate the return value against the
+                output schema. Raises ValidationError if validation fails.
+            **kwargs: Arguments to pass to the tool function.
+
+        Returns:
+            The result of the tool function.
+        """
+        if self.run is None:
+            raise ValueError("No run function provided")
+
+        if self._is_async():
+            result = await self.run(**kwargs)  # type: ignore[func-returns-value]
+        else:
+            loop = asyncio.get_running_loop()
+            assert self.run is not None, "can't run None"
+            result = await loop.run_in_executor(None, lambda: self.run(**kwargs))  # type: ignore
+
+        if validate_output:
+            self._validate_output(result)
+
+        return result
+
+    @classmethod
+    def from_function(
+        cls,
+        func: Callable,
+        *,
+        name: str | None = None,
+        include_output_schema_in_description: bool = False,
+    ) -> "Tool":
+        """
+        Create a Tool from a function using introspection.
+
+        Uses Pydantic's TypeAdapter for robust schema generation, supporting:
+        - All Python types (primitives, generics, unions, Literal, etc.)
+        - Pydantic models and TypedDict as parameter types
+        - Annotated[T, Field(description="...")] for parameter descriptions
+        - Annotated[T, "description"] shorthand for descriptions
+        - Complex nested types with proper $defs/$ref handling
+        - Output schema extraction from return type annotation
+
+        Args:
+            func: The function to create a tool from.
+            name: Optional name override for the tool. If not provided,
+                uses the function's __name__.
+            include_output_schema_in_description: If True, append the return type
+                and any complex type definitions to the tool description. This can
+                help the model understand what the tool returns. Default is False.
+
+        Example:
+            def search(
+                query: Annotated[str, Field(description="Search query")],
+                limit: int = 10,
+                filters: dict[str, str] | None = None,
+            ) -> list[dict]:
+                '''Search the database.'''
+                ...
+
+            tool = Tool.from_function(search)
+            # tool.output_schema contains schema for list[dict]
+            # tool.call(query="test", validate_output=True) validates return value
+
+            # With custom name:
+            tool = Tool.from_function(search, name="search_database")
+            # tool.name is "search_database"
+
+            # With output schema in description:
+            tool = Tool.from_function(search, include_output_schema_in_description=True)
+            # Description becomes:
+            # "Search the database.
+            #
+            # Returns: list[dict]"
+        """
+        # Get function name (use override if provided)
+        tool_name = name if name is not None else func.__name__
+
+        # Get docstring for description
+        description = func.__doc__ or f"Call the {tool_name} function"
+        description = description.strip()
+
+        # Use TypeAdapter for robust schema generation
+        type_adapter = _get_cached_typeadapter(func)
+        schema = type_adapter.json_schema()
+
+        # Clean up the schema (remove titles, additionalProperties: false)
+        schema = _clean_schema(schema)
+
+        # Extract parameters and required from the schema
+        parameters = schema.get("properties", {})
+        required = schema.get("required", [])
+        definitions = schema.get("$defs")
+
+        # Extract output schema from return type annotation
+        output_schema = None
+        output_type_adapter = None
+        sig = inspect.signature(func)
+        return_type = sig.return_annotation
+
+        if return_type is not inspect.Parameter.empty:
+            try:
+                # Resolve string annotations if needed
+                if isinstance(return_type, str):
+                    hints = get_type_hints(func)
+                    return_type = hints.get("return", return_type)
+
+                # Create TypeAdapter for output validation
+                output_type_adapter = TypeAdapter(return_type)
+                output_schema = _clean_schema(output_type_adapter.json_schema())
+            except Exception:
+                # If we can't create a schema for the return type, that's fine
+                # (e.g., for non-serializable types like custom classes)
+                pass
+
+        # Optionally append output schema info to description
+        if (
+            include_output_schema_in_description
+            and return_type is not inspect.Parameter.empty
+        ):
+            output_info = _format_output_schema_for_description(
+                return_type, output_schema
+            )
+            if output_info:
+                description = f"{description}\n\n{output_info}"
+
+        tool = cls(
+            name=tool_name,
+            description=description,
+            parameters=parameters,
+            required=required,
+            definitions=definitions,
+            output_schema=output_schema,
+            run=func,
+        )
+        # Store the TypeAdapter for runtime validation (not serialized)
+        tool._output_type_adapter = output_type_adapter
+        return tool
+
+    @classmethod
+    async def from_mcp_config(
+        cls,
+        config: dict[str, Any],
+        *,
+        timeout: float | None = None,
+    ) -> list["Tool"]:
+        """
+        config: full Claude-Desktop-style dict *or* just its "mcpServers" block
+        Returns {server_key: [Tool, …], …}
+        """
+        # allow caller to pass either the whole desktop file or just the sub-dict
+        servers_block = config.get("mcpServers", config)
+
+        # FastMCP understands the whole config dict directly
+        client = Client({"mcpServers": servers_block}, timeout=timeout)
+        async with client:
+            all_tools = await _load_all_mcp_tools(client)
+
+        # bucket by prefix that FastMCP added (serverkey_toolname)
+        return all_tools
+
+    @classmethod
+    async def from_mcp(
+        cls,
+        server_name: str,
+        *,
+        tool_name: str | None = None,
+        timeout: float | None = None,
+        **server_spec,  # url="…"  OR  command="…" args=[…]
+    ) -> Any:  # Tool | list[Tool]
+        """
+        Thin wrapper for one server.  Example uses:
+
+            Tool.from_mcp(url="https://weather.example.com/mcp")
+            Tool.from_mcp(command="python", args=["./assistant.py"], tool_name="answer_question")
+        """
+        # ensure at least one of command or url is defined
+        if not (server_spec.get("url") or server_spec.get("command")):
+            raise ValueError("most provide url or command")
+        # build a one-server desktop-style dict
+        cfg = {server_name: server_spec}
+        tools = await cls.from_mcp_config(cfg, timeout=timeout)
+        if tool_name is None:
+            return tools
+        for t in tools:
+            if t.name.endswith(f"{tool_name}"):  # prefixed by FastMCP
+                return t
+        raise ValueError(f"Tool '{tool_name}' not found on that server")
+
+    @staticmethod
+    def _tool_from_meta(meta: dict[str, Any], runner) -> "Tool":
+        props = meta["inputSchema"].get("properties", {})
+        req = meta["inputSchema"].get("required", [])
+        addl = meta["inputSchema"].get("additionalProperties")
+        return Tool(
+            name=meta["name"],
+            description=meta.get("description", ""),
+            parameters=props,
+            required=req,
+            additionalProperties=addl,
+            run=runner,
+        )
+
+    def _is_strict_mode_compatible(self) -> bool:
+        """
+        Check if this tool's schema is compatible with OpenAI strict mode.
+        Strict mode requires all objects to have defined properties.
+        """
+
+        def has_undefined_objects(schema: dict | list | Any) -> bool:
+            """Recursively check for objects without defined properties."""
+            if isinstance(schema, dict):
+                # Check if this is an object type without properties
+                if schema.get("type") == "object":
+                    # If additionalProperties is True or properties is missing/empty
+                    if schema.get("additionalProperties") is True:
+                        return True
+                    if "properties" not in schema or not schema["properties"]:
+                        return True
+                # Recursively check nested schemas
+                for value in schema.values():
+                    if has_undefined_objects(value):
+                        return True
+            elif isinstance(schema, list):
+                for item in schema:
+                    if has_undefined_objects(item):
+                        return True
+            return False
+
+        return not has_undefined_objects(self.parameters or {})
+
+    def _json_schema(
+        self, include_additional_properties=False, remove_defaults=False
+    ) -> dict[str, Any]:
+        def _add_additional_properties_recursive(
+            schema: dict | list | Any, remove_defaults: bool = False
+        ) -> dict | list | Any:
+            """Recursively add additionalProperties: false to all object-type schemas.
+            In strict mode (when remove_defaults=True), also makes all properties required."""
+            if isinstance(schema, dict):
+                # Copy the dictionary to avoid modifying the original
+                new_schema = schema.copy()
+
+                # make sure to label arrays and objects
+                if "type" not in new_schema:
+                    if "properties" in new_schema:
+                        new_schema["type"] = "object"
+                    elif "items" in new_schema:
+                        new_schema["type"] = "array"
+
+                # If this is an object type schema, set additionalProperties: false
+                if new_schema.get("type") == "object":
+                    new_schema["additionalProperties"] = False
+
+                    # In strict mode, all properties must be required
+                    if remove_defaults and "properties" in new_schema:
+                        new_schema["required"] = list(new_schema["properties"].keys())
+
+                # Remove default values if requested (for strict mode)
+                if remove_defaults and "default" in new_schema:
+                    del new_schema["default"]
+
+                # Recursively process all values in the dictionary
+                for key, value in new_schema.items():
+                    new_schema[key] = _add_additional_properties_recursive(
+                        value, remove_defaults
+                    )
+
+                return new_schema
+            elif isinstance(schema, list):
+                # Recursively process all items in the list
+                return [
+                    _add_additional_properties_recursive(item, remove_defaults)
+                    for item in schema
+                ]
+            else:
+                # Return primitive values as-is
+                return schema
+
+        # Start with the base schema structure
+        if include_additional_properties and self.parameters:
+            # Apply recursive additionalProperties processing to parameters
+            processed_parameters = _add_additional_properties_recursive(
+                self.parameters, remove_defaults
+            )
+        else:
+            processed_parameters = self.parameters
+
+        # Process definitions too
+        if self.definitions and include_additional_properties:
+            processed_definitions = _add_additional_properties_recursive(
+                self.definitions, remove_defaults
+            )
+        else:
+            processed_definitions = self.definitions
+
+        res = {
+            "type": "object",
+            "properties": processed_parameters,
+            "required": self.required,  # Use the tool's actual required list
+        }
+
+        if include_additional_properties:
+            res["additionalProperties"] = False
+
+        # Include definitions if present (for $ref support)
+        if processed_definitions:
+            res["$defs"] = processed_definitions
+
+        return res
+
+    # ---------- dumpers ----------
+    def for_openai_completions(
+        self, *, strict: bool = True, **kwargs
+    ) -> dict[str, Any]:
+        if self.is_built_in:
+            return {"type": self.type, **self.built_in_args, **kwargs}
+
+        # Check if schema is compatible with strict mode
+        if strict and not self._is_strict_mode_compatible():
+            strict = False
+
+        if strict:
+            # For strict mode, remove defaults and make all parameters required
+            schema = self._json_schema(
+                include_additional_properties=True, remove_defaults=True
+            )
+            schema["required"] = list(
+                (self.parameters or {}).keys()
+            )  # All parameters required in strict mode
+        else:
+            # For non-strict mode, use the original required list
+            schema = self._json_schema(include_additional_properties=True)
+
+        return {
+            "type": "function",
+            "function": {
+                "name": self.name,
+                "description": self.description,
+                "parameters": schema,
+                "strict": strict,
+            },
+        }
+
+    def for_openai(self, strict: bool = True, **kwargs):
+        """just an alias for the above"""
+        return self.for_openai_completions(strict=strict, **kwargs)
+
+    def for_openai_responses(self, *, strict: bool = True, **kwargs) -> dict[str, Any]:
+        if self.is_built_in:
+            return {"type": self.type, **self.built_in_args, **kwargs}
+
+        # Check if schema is compatible with strict mode
+        if strict and not self._is_strict_mode_compatible():
+            strict = False
+
+        if strict:
+            # For strict mode, remove defaults and make all parameters required
+            schema = self._json_schema(
+                include_additional_properties=True, remove_defaults=True
+            )
+            schema["required"] = list(
+                (self.parameters or {}).keys()
+            )  # All parameters required in strict mode
+
+            return {
+                "type": "function",
+                "name": self.name,
+                "description": self.description,
+                "parameters": schema,
+                "strict": True,
+            }
+        else:
+            # For non-strict mode, use the original required list
+            return {
+                "type": "function",
+                "name": self.name,
+                "description": self.description,
+                "parameters": self._json_schema(include_additional_properties=True),
+            }
+
+    def for_anthropic(self, *, strict: bool = True, **kwargs) -> dict[str, Any]:
+        # built-in tools have "name", "type", maybe metadata
+        if self.is_built_in:
+            return {
+                "name": self.name,
+                "type": self.type,
+                **self.built_in_args,
+                **kwargs,
+            }
+
+        # Check if schema is compatible with strict mode
+        if strict and not self._is_strict_mode_compatible():
+            strict = False
+
+        if strict:
+            # For strict mode, remove defaults and make all parameters required
+            schema = self._json_schema(
+                include_additional_properties=True, remove_defaults=True
+            )
+            schema["required"] = list(
+                (self.parameters or {}).keys()
+            )  # All parameters required in strict mode
+
+            return {
+                "name": self.name,
+                "description": self.description,
+                "input_schema": schema,
+                "strict": True,
+            }
+        else:
+            # For non-strict mode, use the original required list
+            return {
+                "name": self.name,
+                "description": self.description,
+                "input_schema": self._json_schema(),
+            }
+
+    def for_google(self) -> dict[str, Any]:
+        """
+        Shape used by google.genai docs.
+        """
+
+        def _strip_additional_properties(schema: Any) -> Any:
+            if isinstance(schema, dict):
+                return {
+                    key: _strip_additional_properties(value)
+                    for key, value in schema.items()
+                    if key != "additionalProperties"
+                }
+            if isinstance(schema, list):
+                return [_strip_additional_properties(item) for item in schema]
+            return schema
+
+        return {
+            "name": self.name,
+            "description": self.description,
+            "parameters": _strip_additional_properties(self._json_schema()),
+        }
+
+    def for_mistral(self) -> dict[str, Any]:
+        return self.for_openai_completions()
+
+    def dump_for(
+        self,
+        provider: Literal[
+            "openai-responses", "openai-completions", "anthropic", "google"
+        ],
+        **kw,
+    ) -> dict[str, Any]:
+        if provider == "openai-responses":
+            return self.for_openai_responses(**kw)
+        if provider == "openai-completions":
+            return self.for_openai_completions(**kw)
+        if provider == "anthropic":
+            return self.for_anthropic(**kw)
+        if provider == "google":
+            return self.for_google()
+        raise ValueError(provider)
+
+    @classmethod
+    def built_in(cls, name: str, **kwargs):
+        if "type" in kwargs:
+            type = kwargs.pop("type")
+        else:
+            type = name
+        return cls(name=name, type=type, is_built_in=True, built_in_args=kwargs)
+
+
+class OpenAIMCPSpec(TypedDict):
+    type: str
+    server_label: str
+    server_url: str
+    headers: dict | None
+    require_approval: str
+
+
+class MCPServer(BaseModel):
+    """
+    Allow MCPServers to be passed directly, if provider supports it.
+    Provider can directly call MCP instead of handling it client-side.
+    Should work with Anthropic MCP connector and OpenAI responses API.
+    """
+
+    name: str
+    url: str
+    # anthropic-specific
+    token: str | None = None
+    configuration: dict | None = None
+    # openai-specific
+    headers: dict | None = None
+
+    # tools cache
+    _tools: list[Tool] | None = None
+
+    @classmethod
+    def from_openai(cls, spec: OpenAIMCPSpec):
+        return cls(
+            name=spec["server_label"],
+            url=spec["server_url"],
+            headers=spec.get("headers"),
+        )
+
+    def for_openai_responses(self):
+        res: dict[str, Any] = {
+            "type": "mcp",
+            "server_label": self.name,
+            "server_url": self.url,
+            "require_approval": "never",
+        }
+        if self.headers:
+            res["headers"] = self.headers
+
+        return res
+
+    def for_anthropic(self):
+        res: dict[str, Any] = {
+            "type": "url",
+            "url": self.url,
+            "name": self.name,
+        }
+        if self.token:
+            res["authorization_token"] = self.token
+        if self.configuration:
+            res["tool_configuration"] = self.configuration
+
+        return res
+
+    async def to_tools(self) -> list[Tool]:
+        """
+        Compatible with ALL providers.
+        Caches so we don't have to hit the server a ton of times.
+        """
+        if self._tools:
+            return self._tools
+        else:
+            tools: list[Tool] = await Tool.from_mcp(self.name, url=self.url)
+            self._tools = tools
+            return tools
+
+
+# Note: prefab submodule is available via lm_deluge.tool.prefab
+# but not auto-imported here to avoid circular imports