pygent-ai 0.1.5

A Python framework for building LLM-powered agents with modular state, tools, and MCP support.

Pygent

A Python framework for building LLM-powered agents with modular state management, composable operators, and LLM tool-calling integration (including MCP).

Features

• Modular architecture — Agents, context, tools, memory, and plans as composable modules
• State management — Consistent save/load and serialization via PygentOperator
• LLM-native — Messages and tools align with OpenAI-style APIs
• MCP support — Use Model Context Protocol (SSE and stdio) tools via ToolManager
• Toolkits — Built-in toolkits such as TerminalToolkits for shell access; RestrictedTerminal remains as a compatibility alias.

Requirements

• Python 3.11+

Installation

From the project root:

pip install -e .

Or with uv:

uv pip install -e .

Quick Start

from pygent.agent import BaseAgent
from pygent.context import BaseContext
from pygent.llm import AsyncRequestsClient
from pygent.message import UserMessage
from pygent.module.tool import ToolManager
from pygent.toolkits import TerminalToolkits

class MyAgent(BaseAgent):
    def __init__(self):
        super().__init__()
        self.llm = AsyncRequestsClient(
            base_url="https://api.openai.com/v1",
            api_key="YOUR_API_KEY",
            model_name="gpt-4",
        )
        self.tool_manager = ToolManager()
        terminal = TerminalToolkits(session_id="quickstart", workspace_root=".")
        self.tool_manager.add_module("terminal", terminal)
        self.tool_manager.register_tools(terminal.get_all_tools())

    async def run(self, user_input: str):
        context = BaseContext(system_prompt="You are a helpful assistant.")
        context.add_message(UserMessage(content=user_input))
        # ... tool loop and LLM calls (see examples/react_agent)
        return context

See examples/react_agent/react_agent.py for a full ReAct-style agent with tool calling.

Compatibility note: AsyncOpenAIClient is still exported as an alias of AsyncRequestsClient, and RestrictedTerminal(root_dir=...) still delegates to TerminalToolkits.

Project Layout

• pygent/ — Core package: agents, context, LLM client, messages, modules (tool, plan, memory, etc.), toolkits
• examples/ — Example agents (e.g. ReAct agent)
• MCPs/ — Example MCP servers (e.g. LocalMemoryMCP)
• docs/ — Architecture and API documentation
• tests/ — Pytest tests

Documentation

• Architecture — Design, layers, and components
• API — API reference

License

Apache-2.0. See LICENSE for details.