polytools

Cross-provider LLM tool schema generation from Python type hints.

Write your tool once. Export to OpenAI, Anthropic, Gemini, and MCP — no framework lock-in, no third-party dependencies.

from polytools import tool

@tool
def search_web(query: str, max_results: int = 5) -> list[str]:
    """Search the web and return relevant URLs.

    Args:
        query: The search query string.
        max_results: Maximum number of results to return.
    """
    ...

search_web.to_openai()     # → OpenAI function calling format
search_web.to_anthropic()  # → Anthropic tools format
search_web.to_gemini()     # → Gemini FunctionDeclaration format
search_web.to_mcp()        # → MCP JSON-RPC tool definition
search_web.to_all()        # → all four at once

Why polytools?

Every LLM provider uses a different JSON format for tool/function definitions. Today you either hand-write four separate schemas, lock into a framework that weighs megabytes, or wrestle Pydantic into every project.

polytools is a single decorator — pure Python stdlib, zero external dependencies — that reads your type hints and docstring once and outputs whatever format you need.

Installation

pip install polytools

Requires Python 3.9+. No other dependencies, ever.

Quick Start

from polytools import tool
from typing import Optional, Literal

@tool
def send_email(
    recipient: str,
    subject: str,
    body: str,
    cc: Optional[str] = None,
    priority: Literal["low", "normal", "high"] = "normal",
) -> bool:
    """Send an email message.

    Args:
        recipient: Email address of the primary recipient.
        subject: Subject line of the email.
        body: Full text body of the email.
        cc: Optional CC email address.
        priority: Message priority level.
    """
    ...

OpenAI

import openai

client = openai.OpenAI()
response = client.chat.completions.create(
    model="gpt-4o",
    messages=[{"role": "user", "content": "Send an email to [email protected]"}],
    tools=[send_email.to_openai()],
)

Schema output

{
  "type": "function",
  "function": {
    "name": "send_email",
    "description": "Send an email message.",
    "parameters": {
      "type": "object",
      "properties": {
        "recipient": {"type": "string", "description": "Email address of the primary recipient."},
        "subject":   {"type": "string", "description": "Subject line of the email."},
        "body":      {"type": "string", "description": "Full text body of the email."},
        "cc":        {"anyOf": [{"type": "string"}, {"type": "null"}], "description": "Optional CC email address."},
        "priority":  {"type": "string", "enum": ["low", "normal", "high"], "description": "Message priority level."}
      },
      "required": ["recipient", "subject", "body"]
    }
  }
}

Anthropic

import anthropic

client = anthropic.Anthropic()
response = client.messages.create(
    model="claude-sonnet-5",
    messages=[{"role": "user", "content": "Send an email to [email protected]"}],
    tools=[send_email.to_anthropic()],
)

Schema output

{
  "name": "send_email",
  "description": "Send an email message.",
  "input_schema": {
    "type": "object",
    "properties": {
      "recipient": {"type": "string", "description": "Email address of the primary recipient."},
      "subject":   {"type": "string", "description": "Subject line of the email."},
      "body":      {"type": "string", "description": "Full text body of the email."},
      "cc":        {"anyOf": [{"type": "string"}, {"type": "null"}], "description": "Optional CC email address."},
      "priority":  {"type": "string", "enum": ["low", "normal", "high"], "description": "Message priority level."}
    },
    "required": ["recipient", "subject", "body"]
  }
}

Gemini

import google.generativeai as genai

model = genai.GenerativeModel(
    model_name="gemini-1.5-pro",
    tools=[{"function_declarations": [send_email.to_gemini()]}],
)

Schema output

{
  "name": "send_email",
  "description": "Send an email message.",
  "parameters": {
    "type": "OBJECT",
    "properties": {
      "recipient": {"type": "STRING", "description": "Email address of the primary recipient."},
      "subject":   {"type": "STRING", "description": "Subject line of the email."},
      "body":      {"type": "STRING", "description": "Full text body of the email."},
      "cc":        {"type": "STRING", "description": "Optional CC email address."},
      "priority":  {"type": "STRING", "enum": ["low", "normal", "high"], "description": "Message priority level."}
    },
    "required": ["recipient", "subject", "body"]
  }
}

MCP

# In your MCP server's tools/list handler:
tools = [send_email.to_mcp()]

Schema output

{
  "name": "send_email",
  "description": "Send an email message.",
  "inputSchema": {
    "type": "object",
    "properties": {
      "recipient": {"type": "string", "description": "Email address of the primary recipient."},
      "subject":   {"type": "string", "description": "Subject line of the email."},
      "body":      {"type": "string", "description": "Full text body of the email."},
      "cc":        {"anyOf": [{"type": "string"}, {"type": "null"}], "description": "Optional CC email address."},
      "priority":  {"type": "string", "enum": ["low", "normal", "high"], "description": "Message priority level."}
    },
    "required": ["recipient", "subject", "body"]
  }
}

Invoking from an LLM response

import json

# OpenAI
tool_call = response.choices[0].message.tool_calls[0]
result = send_email.call(json.loads(tool_call.function.arguments))

# Anthropic
block = response.content[0]  # ToolUseBlock
result = send_email.call(block.input)

Supported Type Annotations

Python type	JSON Schema output
str	{"type": "string"}
int	{"type": "integer"}
float	{"type": "number"}
bool	{"type": "boolean"}
bytes	{"type": "string", "format": "byte"}
None / type(None)	{"type": "null"}
list / List[T] / list[T]	{"type": "array", "items": {...}}
dict / Dict[K, V] / dict[K, V]	{"type": "object", "additionalProperties": {...}}
tuple[T, ...]	{"type": "array", "items": {...}}
tuple[T1, T2, T3]	{"type": "array", "prefixItems": [...]}
set[T] / frozenset[T]	{"type": "array", "uniqueItems": true, "items": {...}}
Optional[T]	{"anyOf": [{...}, {"type": "null"}]}
Union[T1, T2]	{"anyOf": [{...}, {...}]}
Literal["a", "b"]	{"type": "string", "enum": ["a", "b"]}
Enum subclass	{"type": <value type>, "enum": [...]}
dataclass	{"type": "object", "properties": {...}, "required": [...]}
TypedDict	{"type": "object", "properties": {...}, "required": [...]}
NamedTuple	{"type": "object", "properties": {...}, "required": [...]}
Any	{} (no constraints)
Unannotated	{} (no constraints)

Nested types are fully supported: list[dict[str, Optional[int]]] works as expected.

Structured inputs

Parameters typed as a dataclass, TypedDict, NamedTuple, or Enum become proper nested JSON Schema objects — no Pydantic required. A field is marked required unless it has a default (or is a total=False / NotRequired TypedDict key).

from dataclasses import dataclass
from enum import Enum
from polytools import tool

class Priority(Enum):
    LOW = "low"
    HIGH = "high"

@dataclass
class Ticket:
    title: str
    priority: Priority
    assignee: str = ""

@tool
def create_ticket(ticket: Ticket) -> str:
    """Open a support ticket.

    Args:
        ticket: The ticket to create.
    """
    ...

create_ticket.to_openai()

Schema output (parameters)

{
  "type": "object",
  "properties": {
    "ticket": {
      "type": "object",
      "description": "The ticket to create.",
      "properties": {
        "title":    {"type": "string"},
        "priority": {"type": "string", "enum": ["low", "high"]},
        "assignee": {"type": "string"}
      },
      "required": ["title", "priority"]
    }
  },
  "required": ["ticket"]
}

Nesting is recursive (a dataclass field that is itself a dataclass, list[SomeDataclass], Optional[SomeTypedDict], dict[str, SomeEnum], …). Self-referential types are guarded and collapse to a bare {"type": "object"}.

Note: .call(args) passes the LLM's argument dict straight to your function; it does not reconstruct dataclass/TypedDict instances from that dict. Schema generation and argument coercion are separate concerns — coercion may land in a future release.

Docstring Styles

Parameter descriptions are parsed automatically from Google, NumPy, and reStructuredText docstrings:

Google (recommended)

def f(x: int, y: str = "hello") -> bool:
    """Summary line.

    Args:
        x: Description of x.
        y: Description of y.
    """

NumPy

def f(x: int, y: str = "hello") -> bool:
    """Summary line.

    Parameters
    ----------
    x : int
        Description of x.
    y : str, optional
        Description of y.
    """

reStructuredText

def f(x: int, y: str = "hello") -> bool:
    """Summary line.

    :param x: Description of x.
    :param y: Description of y.
    """

API Reference

@tool

Decorator that wraps a Python function and exposes provider schema methods. Can be used with or without parentheses:

@tool
def my_func(...): ...

@tool()
def my_func(...): ...

Tool methods

Method	Returns	Description
.to_openai()	dict	OpenAI Chat Completions tools list entry
.to_anthropic()	dict	Anthropic Messages API tools list entry
.to_gemini()	dict	Gemini function_declarations list entry
.to_mcp()	dict	MCP tools/list response entry
.to_all()	dict[str, dict]	All four, keyed by provider name
.call(args: dict)	Any	Invoke the function with an LLM argument dict

Tool attributes

Attribute	Type	Description
__wrapped__	Callable	The original unwrapped function
__name__	str	Preserved from the original function
__doc__	str	Preserved from the original function

Notes on Gemini

Gemini's API does not support anyOf in tool schemas. Optional[T] parameters are handled by omitting them from the required list rather than encoding nullability in the schema type. Union types are resolved to the first non-null type.

Contributing

See CONTRIBUTING.md.

License

MIT — see LICENSE.

---

Part of the aenealabs AI agent toolkit.