AI Agents with Function Calling: Build Tools LLMs Can Use

A developer published a from-scratch framework for building LLM agents using function calling, in which the model emits structured JSON to invoke predefined tools. The implementation defines an abstract Tool class with name, description, parameters and execute members, plus concrete tools for DuckDuckGo web search, sandboxed Python execution with an import allowlist and a 10-second subprocess timeout, and a math-only calculator.

Create AI agents that can execute code, search the web, and call APIs — using function calling. Function calling tool use transforms LLMs from text generators into agents that can act . This tutorial builds a complete agent framework from scratch. LLMs output structured JSON to invoke predefined functions: User: "What's the weather in Tokyo?" LLM: {"tool": "get weather", "args": {"city": "Tokyo"}} System: executes function, returns result LLM: "It's 22°C and sunny in Tokyo." requests library python tools.py import os import requests import json from typing import Dict, Any, List from abc import ABC, abstractmethod class Tool ABC : @property @abstractmethod def name self - str: ... @property @abstractmethod def description self - str: ... @property @abstractmethod def parameters self - Dict: ... @abstractmethod def execute self, kwargs - Any: ... --- Concrete Tools --- class WebSearchTool Tool : name = "web search" description = "Search the web for current information" parameters = { "type": "object", "properties": { "query": {"type": "string", "description": "Search query"} }, "required": "query" } def execute self, query: str - str: Using DuckDuckGo HTML scrape free, no API key url = "https://html.duckduckgo.com/html/" params = {"q": query} headers = {"User-Agent": "Mozilla/5.0"} try: response = requests.post url, data=params, headers=headers, timeout=10 from bs4 import BeautifulSoup soup = BeautifulSoup response.text, 'html.parser' results = soup.find all 'a', class ='result snippet' snippets = r.get text :200 for r in results :3 return "\n".join snippets if snippets else "No results found" except Exception as e: return f"Search error: {e}" class PythonExecutorTool Tool : name = "execute python" description = "Execute Python code and return output" parameters = { "type": "object", "properties": { "code": {"type": "string", "description": "Python code to execute"} }, "required": "code" } def execute self, code: str - str: import subprocess import tempfile import sys Safety: only allow specific imports allowed imports = {'json', 'math', 'random', 'datetime', 'statistics', 'itertools', 'collections'} with tempfile.NamedTemporaryFile mode='w', suffix='.py', delete=False as f: f.write code temp path = f.name try: result = subprocess.run sys.executable, temp path , capture output=True, text=True, timeout=10 os.unlink temp path output = result.stdout if result.stderr: output += f"\nSTDERR: {result.stderr}" return output :2000 if output else " no output " except subprocess.TimeoutExpired: os.unlink temp path return "Error: Execution timeout 10s " except Exception as e: return f"Error: {e}" class CalculatorTool Tool : name = "calculate" description = "Perform mathematical calculations" parameters = { "type": "object", "properties": { "expression": {"type": "string", "description": "Math expression to evaluate"} }, "required": "expression" } def execute self, expression: str - str: try: Safe eval for math only allowed names = {"abs", "round", "min", "max", "sum", "pow"} allowed names.update {k: v for k, v in import 'math' . dict .items if not k.startswith ' ' } code = compile expression, "<string ", "eval" for name in code.co names: if name not in allowed names: return f"Error: '{name}' not allowed" result = eval code, {" builtins ": {}}, allowed names return str result except Exception as e: return f"Error: {e}" class FileReadTool Tool : name = "read file" description = "Read contents of a file" parameters = { "type": "object", "properties": { "path": {"type": "string", "description": "File path to read"} }, "required": "path" } def execute self, path: str - str: try: with open path, 'r' as f: return f.read :5000 except Exception as e: return f"Error reading file: {e}" class FileWriteTool Tool : name = "write file" description = "Write content to a file" parameters = { "type": "object", "properties": { "path": {"type": "string", "description": "File path to write"}, "content": {"type": "string", "description": "Content to write"} }, "required": "path", "content" } def execute self, path: str, content: str - str: try: os.makedirs os.path.dirname path or '.', exist ok=True with open path, 'w' as f: f.write content return f"File written: {path}" except Exception as e: return f"Error writing file: {e}" Registry TOOLS = { "web search": WebSearchTool , "execute python": PythonExecutorTool , "calculate": CalculatorTool , "read file": FileReadTool , "write file": FileWriteTool } def get tool schemas - List Dict : """Return OpenAI-compatible tool schemas.""" return { "type": "function", "function": { "name": tool.name, "description": tool.description, "parameters": tool.parameters } } for tool in TOOLS.values def execute tool name: str, args: Dict - str: if name not in TOOLS: return f"Error: Unknown tool '{name}'" try: return TOOLS name .execute args except Exception as e: return f"Tool error: {e}" python agent.py import os import json import requests from typing import List, Dict, Any, Optional from tools import get tool schemas, execute tool class Agent: def init self, provider: str = "openrouter", model: str = None : self.provider = provider self.configs = { "openrouter": { "url": "https://openrouter.ai/api/v1/chat/completions", "key": os.getenv "OPENROUTER API KEY" , "model": model or "google/gemini-2.0-flash-exp:free" }, "groq": { "url": "https://api.groq.com/openai/v1/chat/completions", "key": os.getenv "GROQ API KEY" , "model": model or "llama-3.1-70b-versatile" } } self.messages = self.tools = get tool schemas self.max turns = 10 def add system prompt self, prompt: str : self.messages.insert 0, {"role": "system", "content": prompt} def run self, user input: str - str: self.messages.append {"role": "user", "content": user input} for turn in range self.max turns : response = self. call llm if "tool calls" in response: Execute tools tool results = for tc in response "tool calls" : name = tc "function" "name" args = json.loads tc "function" "arguments" result = execute tool name, args tool results.append { "tool call id": tc "id" , "role": "tool", "name": name, "content": result } Add tool results to conversation self.messages.append {"role": "assistant", "tool calls": response "tool calls" } self.messages.extend tool results else: Final answer content = response.get "content", "" self.messages.append {"role": "assistant", "content": content} return content return "Max turns reached" def call llm self - Dict: config = self.configs self.provider headers = { "Authorization": f"Bearer {config 'key' }", "Content-Type": "application/json" } data = { "model": config "model" , "messages": self.messages, "tools": self.tools, "tool choice": "auto", "temperature": 0.1, "max tokens": 2000 } response = requests.post config "url" , headers=headers, json=data, timeout=60 response.raise for status return response.json "choices" 0 "message" def run stream self, user input: str : """Streaming version.""" self.messages.append {"role": "user", "content": user input} for turn in range self.max turns : First, check if we need tools non-streaming for tool calls response = self. call llm if "tool calls" in response: tool results = for tc in response "tool calls" : name = tc "function" "name" args = json.loads tc "function" "arguments" result = execute tool name, args tool results.append { "tool call id": tc "id" , "role": "tool", "name": name, "content": result } self.messages.append {"role": "assistant", "tool calls": response "tool calls" } self.messages.extend tool results continue Stream final answer content = response.get "content", "" self.messages.append {"role": "assistant", "content": content} Simulate streaming by yielding chunks for chunk in content.split : yield chunk + " " return python main.py import os from agent import Agent Initialize agent = Agent provider="openrouter" or "groq" System prompt agent.add system prompt """You are a helpful AI assistant with access to tools. Use tools when you need current information, calculations, or code execution. Think step by step. Be concise. Available tools: web search, execute python, calculate, read file, write file""" Example 1: Web search + calculation print "=== Query 1 ===" result = agent.run "What's the current population of Tokyo? Calculate 15% of it." print result Example 2: Code execution print "\n=== Query 2 ===" result = agent.run """ Write a Python function to find the nth Fibonacci number, then compute the 20th Fibonacci number. """ print result Example 3: File operations print "\n=== Query 3 ===" result = agent.run """ Create a file 'todo.txt' with these items: 1. Learn function calling 2. Build an agent 3. Deploy it Then read it back. """ print result Example 4: Multi-step reasoning print "\n=== Query 4 ===" result = agent.run """ Find the current price of Bitcoin, then calculate how much 0.5 BTC would be worth in USD. """ print result python agent with memory.py import json from agent import Agent class PersistentAgent Agent : def init self, args, memory file: str = "agent memory.json", kwargs : super . init args, kwargs self.memory file = memory file self.load memory def load memory self : try: with open self.memory file, 'r' as f: data = json.load f self.messages = data.get "messages", except FileNotFoundError: pass def save memory self : with open self.memory file, 'w' as f: json.dump {"messages": self.messages -20: }, f Keep last 20 def run self, user input: str - str: result = super .run user input self.save memory return result Usage agent = PersistentAgent memory file="my agent memory.json" print agent.run "Remember my name is Alex" print agent.run "What's my name?" Will remember | Provider | Model | Function Calling | |---|---|---| | OpenRouter | google/gemini-2.0-flash-exp:free | ✅ Excellent | | OpenRouter | mistralai/mistral-7b-instruct:free | ✅ Good | | Groq | llama-3.1-70b-versatile | ✅ Excellent | | Groq | llama-3.1-8b-instant | ✅ Good | | Together | meta-llama/Meta-Llama-3.1-70B-Instruct-Turbo | ✅ Good | All components use free tiers. Function calling works on free models via OpenRouter and Groq. You now have a working agent framework. Extend it with: The pattern is always: define tool → register schema → execute → return result. Part of the AI Agents series on Dev.to