VenusAI - Advanced Agent Framework 🚀

VenusAI is a secure and extensible Agent framework built for modern AI applications. It offers dynamic tool management, powerful decorators, advanced caching, robust error handling, a built-in CLI, and seamless Claude MCP integration.

---

🔑 Key Capabilities

• 🛡️ Security-first design with permission-based tool filtering & E2B sandbox integration
• 🔧 Dynamic tool ecosystem with decorators for safety, autofix & error recovery
• ⚡ High-performance caching with multiple backends (aiocache, lrucache, async-lru, cachetools)
• 🌐 HTTP API generation → automatically expose tools as REST endpoints
• 🤖 MCP Protocol native support → seamless Claude Desktop integration
• 🎯 Type-safe dependency injection with advanced context management
• 🔄 Self-healing tools → automatic error recovery & retry mechanisms
• 📊 Comprehensive error tracking with detailed frame info & custom handlers

Whether you're building simple chatbots or complex multi-agent systems, VenusAI provides the foundation for scalable, maintainable, and secure AI applications.

---

✨ Features

🔹 Core Bases

• Venus

  • Base class for all Agents
  • No default toolset (bare Agent)

• VenusCode

  • Subclass of Venus with coding capabilities
  • Built-in filesystem toolset
  • Permission-based tool filtering (supports custom permitters)
  • Code execution disabled by default
  • E2B sandbox integration for safe execution

---

🔹 Tools

• Dynamic tool integration from modules

• Dynamic Dependency Injection

• Decorators

  • @agent.safe → error-safe wrapper for non-context tools
  • @agent.safe_plain → error-safe wrapper for context tools
  • @agent.autofix → self-healing tools (functions can fix themselves)
  • @agent.on_error → custom error handler

• Register tools as HTTP endpoints (beta)

  • Convert registered tools to HTTP API (via FastAPI)
  • Just call agent.tools_http_api() and agent.serve()

• Sync/Async caching for tools with @cached

  • Backends: aiocache, lrucache, async-lru, cachetools

• Autofix mechanism

  • Implicitly handles errors via @safe
  • Customizable fix-prompt & fix-model
  • Falls back to a default model if none provided

• Error Handlers

  • Errors yield an ErrorDict with frame & error details
  • Fully customizable responses/actions

---

🔹 Example

from venus import Venus
from venus.errors import ErrorDict
from venus.types import Deps, DepsT, ModelRetry, RunContext
import hashlib

agent = Venus("grok:grok-3", deps_type=int)

class Bank(Deps[DepsT]):
    reserve: int
    """Current bank reserves."""

@agent.on_error
async def retry_on_failure(err: ErrorDict):
    print(f"Error occurred: {err.exception} at {err.location}. Retrying...")
    raise ModelRetry(err.exception)

@agent.on_error
async def notify(err: ErrorDict):
    # e.g: await send_mail(body=err.message)
    pass

def get_reserves():
    return 1_881_938

def get_details():
    return {'code': 'tr', 'swift': 1283, 'id': 1710}

@agent.safe(retries=3, deps=Deps.new(reserve=get_reserves, details=get_details))
async def add_money(ctx: RunContext[Bank[int]], fund: int):
    if fund <= 5:
        raise ValueError("Enter a number greater than 5.")
    
    ctx.deps.reserve += fund
    bank_details = ctx.deps.get(dict)
    bank_id = bank_details['id']
    tx_hash = hashlib.md5(str(bank_id + ctx.deps.reserve).encode()).hexdigest()
    
    print(f"Connected bank with ID {bank_details['code'].upper()}{bank_details['swift']}")
    print(f"Added ${fund} to current (${ctx.deps.reserve - fund}) reserves.")
    print(f"Hash for transaction: {tx_hash}")
    
    return ctx.deps.reserve

Run:

result = agent.run_sync("Add random money to the bank, pick 4 to 6.", output_type=int)
print(result.output)

or

a2a = agent.to_a2a()

venus serve agent:agent a2a --env dev

✅ This example is complete and runnable as-is.

---

🔹 MCP (Model Context Protocol)

• Tool integration from modules via @tool / @mcp_tool
• Dynamic Claude configuration with MCP.configure(configure_claude=True)
• Dependency Injection support for MCP tools
• mcp-remote integration with HTTP/SSE for Claude Desktop

---

🔹 CLI

Venus provides a command-line interface (CLI) to manage and run agents. You can start chats, serve APIs, or launch MCP servers directly from the terminal.

Available Commands

• Chat with an agent

venus chat module:app

• Run MCP Server

venus mcp --path my_tools.py --name "Venus MCP" --host 127.0.0.1 --port 8000 --transport <sse|http|stdio> --configure

venus mcp --path my_tools.py --name "Venus MCP" --host 127.0.0.1 --port 8000 --transport <sse|http|stdio> --configure --all

• Serve an Agent as API

venus serve mymodule:agent --auto --env dev

CLI Options

• chat → Start interactive CLI chat with an agent
• mcp → Run an MCP server with tools from modules
• serve → Expose your agent via HTTP (FastAPI/Uvicorn)
• Supports plugins such as A2A (a2a)

---

⚡ Usage Examples

Basic Agent

from agent import Venus

agent = Venus(name="venus")
response = agent.run_sync("Hello there!")
print(result.output)

Code-Capable Agent

from venus import VenusCode
from venus.permissions import Permissions

code_agent = VenusCode(name="coder", permission=Permissions.READ_EXECUTE)

Dependency Injection

from venus import Venus
from venus.types import Deps, DepsT, RunContext

import uuid
import time

agent = Venus()

uuidgen = lambda: uuid.uuid4().hex
datagen = lambda: {'foo': [Deps(bar='baz')]}

class Auth(Deps[DepsT]):
    id: str

@agent.safe(deps=Deps(id=uuidgen, data=datagen))
def get_tx(ctx: RunContext[Auth]):
    # attribute-style access to deps entity `id`
    txhash = f'%d$%s' % (time.time(), ctx.deps.id)
     # type-based access to deps entity `foo`
    data = ctx.deps.get(dict) # None
    data = ctx.deps.get(list) # [Deps(bar='baz')]
    
    # type-based access to deps entity `foo`
    data = ctx.deps.get(list[Deps]) # [Deps(bar='baz')]
    return txhash + data.bar

Module Tools with Decorators

# agent.py
from venus import Venus
agent = Venus(tool_modules='agent_tools')

# agent_tools.py
from venus.types import Deps
from venus.caching import cached
from venus.decorators import tool, mcp_tool, safe_call, autofix

@tool
@cached(ttl=240)
def get_id():
    return 1

@mcp_tool(deps=Deps(id=get_id))
def get_username(deps: Deps):
    return f'@user{deps.id}'

@safe_call
async def create_user(username: str):
    return True

@autofix
async def risky_function():
    raise Exception('An error occured')

Agent Tools with Decorators

# agent.py
from venus import Venus
from venus.types import RunContext

agent = Venus()

@agent.safe_plain
def add(x: int, y: int) -> int:
    return x + y

@agent.safe(retries=3)
def sub(ctx: RunContext, x: int, y: int) -> int:
    return x - y

@agent.autofix(retries=2, deps=Deps(result=lambda: 20))
def risky_function(data: str):
    raise Exception('An error occured')

---

🛠 Tech Stack

• Python 3.10+ → async-first with modern type hints
• Based on PydanticAI → robust validation & AI agent foundation
• ASGI-compatible → works with FastAPI, Uvicorn, etc.
• MCP Protocol → native Model Context Protocol integration
• Secure execution with E2B Sandbox
• CLI powered by Click → ergonomic, extensible command line
• Advanced Caching → multiple backend support
• Dependency Injection → type-safe, dynamic DI system
• Error Handling → custom recovery & retry strategies
• Decorator System → tool safety, autofix & error control
• HTTP API Generation → auto REST endpoint conversion

---

🤝 Contributing

Contributions are welcome! 🎉 Please open an issue before submitting a PR to discuss your idea.

---

📜 License

Licensed under the MIT License – see the LICENSE file for details.