agentlys 1.15.0

Lightweight AI agent library. Turn Python functions/classes into AI tools instantly.

Agentlys

Turn any Python class into an AI tool. Instantly.

Async-native • MCP support • Multi-providers • ~500 lines of core code

class Database:
    def __llm__(self):
        return f"Tables: {self.list_tables()}"  # AI sees this every turn

    def query(self, sql: str) -> list[dict]:
        """Execute SQL query"""
        return self.execute(sql)

    def describe(self, table: str) -> dict:
        """Get table schema"""
        return self.get_schema(table)

agent = Agentlys()
agent.add_tool(Database(conn))  # That's it. All methods are now AI tools.
agent.run("What drove revenue decline in Q3?")

Other frameworks: 50 lines of tool definitions, separate schemas, manual state management.
Agentlys: Your class IS the tool. Methods become actions. __llm__() injects state.

---

Why Agentlys?

If you want...	Use
Graphs and state machines	LangGraph
Team-based agent crews	CrewAI
Your existing classes as AI tools	Agentlys

~500 lines of core code. No framework lock-in. No magic.

---

Install

pip install 'agentlys[all]'  # OpenAI + Anthropic + MCP

---

The Pattern

1. Functions → Tools

def get_weather(city: str) -> str:
    """Get current weather for a city"""
    return requests.get(f"https://wttr.in/{city}?format=3").text

agent.add_function(get_weather)

2. Classes → Stateful Tools (the killer feature)

class FileSystem:
    def __init__(self, root: str):
        self.root = root

    def __llm__(self):
        """State shown to AI each turn"""
        return f"Current directory: {self.root}\nFiles: {os.listdir(self.root)}"

    def read(self, path: str) -> str:
        """Read file contents"""
        return open(f"{self.root}/{path}").read()

    def write(self, path: str, content: str):
        """Write to file"""
        open(f"{self.root}/{path}", 'w').write(content)

agent.add_tool(FileSystem("/workspace"))
# AI now sees file state, can read/write, all from one class

3. Run Conversations

for message in agent.run_conversation("Refactor config.json to use environment variables"):
    print(message.content)

Async Support

# Async conversation loop
async for message in agent.run_conversation_async("Analyze the data"):
    print(message.content)

# Single async call
response = await agent.ask_async("What tables exist?")

---

Real Example: agentlys-dev

A coding agent in 15 lines:

from agentlys import Agentlys
from agentlys_tools import CodeEditor, Terminal, Git

agent = Agentlys(
    instruction="You are a senior developer",
    provider="anthropic",
    model="claude-sonnet-4-20250514"
)

agent.add_tool(CodeEditor())
agent.add_tool(Terminal())
agent.add_tool(Git())

agent.run_conversation("Create a FastAPI app with tests")

---

Providers

# Anthropic (default)
agent = Agentlys(provider="anthropic", model="claude-sonnet-4-20250514")

# OpenAI
agent = Agentlys(model="gpt-4o")

# Any OpenAI-compatible API (Ollama, vLLM, LiteLLM, OpenRouter, ...)
agent = Agentlys(model="llama3.1", base_url="http://localhost:11434/v1")

See the Provider Guide for details.

---

4. Tool Search (defer loading)

When you have many tools, defer most of them and let the LLM discover what it needs:

agent.add_tool(Database(conn), "db")
agent.add_tool(Charts(), "charts")
agent.add_tool(Documents(), "docs")
agent.add_function(answer)

agent.enable_tool_search(always_loaded=["answer", "Database-db__query"])
# Charts and Documents tools are hidden until the LLM searches for them

Reduces context usage by 50-85% with large tool sets. See API Reference.

---

More

• API Reference
• Sub-Agents
• Examples
• MCP Integration

---

Used By

• Myriade — AI-native data platform

When NOT to use Agentlys

• You need graph-based workflows → Use LangGraph
• You want pre-built agent teams → Use CrewAI
• You need sandboxed code execution → Use Smolagents

Agentlys is for: turning your existing Python code into AI tools with zero ceremony.

License

MIT