Veska

A multi-agent AI framework built from scratch in Python. No LangChain, no CrewAI, no AutoGen — 100% custom-built for full control.

Install

pip install veska

Quick Start

from veska import Agent

agent = Agent(
    name="assistant",
    system_prompt="You are a helpful coding assistant.",
    model="claude-sonnet-4-6",
)

result = agent.run("Explain what a decorator is in Python")
print(result.output)

Add Tools

from veska import Agent, tool

@tool
def get_weather(city: str):
    return f"Weather in {city}: 72°F, sunny"

agent = Agent(
    name="weather-bot",
    system_prompt="You help users check the weather. Use the get_weather tool.",
    model="claude-sonnet-4-6",
    tools=[get_weather],
)

result = agent.run("What's the weather in Paris?")
print(result.output)

Structured Output

from veska import Agent

agent = Agent(
    name="reviewer",
    system_prompt="You review movies.",
    model="claude-sonnet-4-6",
    output_format={
        "title": str,
        "rating": float,
        "recommend": bool,
    }
)

result = agent.run("Review the movie Inception")
print(result.output["title"])      # "Inception"
print(result.output["rating"])     # 9.0

Streaming

result = agent.run("Write a haiku about coding", stream=True)

Multi-Agent System

from veska import Agent, Orchestrator

researcher = Agent(
    name="researcher",
    system_prompt="You research topics thoroughly.",
    model="claude-sonnet-4-6",
)

writer = Agent(
    name="writer",
    system_prompt="You write clear, engaging content.",
    model="claude-sonnet-4-6",
)

orchestrator = Orchestrator(
    model="claude-sonnet-4-6",
    agents=[researcher, writer],
    tools=["file_manager"],
)

result = orchestrator.run("Write a blog post about AI agents")
print(result.results)

Per-Agent Models

researcher = Agent(name="researcher", model="claude-sonnet-4-6")
writer = Agent(name="writer", model="gpt-4o")

Features

• Multi-Agent Orchestration — Orchestrator breaks tasks into a dependency graph, runs them in parallel/sequential order
• Multi-Model Support — Claude and OpenAI, configurable per agent
• Unified Tool System — Pre-built, custom, and MCP tools. Just use @tool decorator
• Streaming — stream=True or stream=callback
• Structured Output — Pass output_format dict, get validated responses
• Memory System — Private memory per agent + shared memory pool
• 3-Level Error Recovery — Auto-retry, agent-level fix, discussion room
• Security Sandboxing — Agents sandboxed to their own territory
• Extended Thinking — Optional per-agent thinking support
• MCP Support — Connect external services via Model Context Protocol
• General Purpose — Not locked to any use case

Requirements

• Python 3.10+
• anthropic, openai, pydantic