pureagents 0.1.0

LLM agents without the complexity. Just agents.

The simplest agent framework. And the most capable.

Documentation · Quick start · API

---

A Python library for building LLM agents. One file you can read in 30 minutes. Tools, streaming, memory, structured outputs, multi-agent orchestration. No magic, no abstractions you don't need.

---

Installation

pip install pureagents

Usage

from pure_agents import Agent, tool

@tool
def search(query: str) -> str:
    """Search the web."""
    return f"Results for {query}..."

agent = Agent(tools=[search])
result = await agent.run("Find the weather in Madrid")

Features

Feature	Usage
Tools	@tool decorator, type hints become JSON schemas
Providers	provider="openai" (Mistral, OpenAI, Anthropic)
Streaming	agent.stream()
Memory	session="my-chat"
Structured outputs	output=MyDataclass
Hooks	on_tool_call=fn
Batch	agent.batch([...])
Chaining	chain([agent1, agent2], prompt)
Routing	Router(agents={...}, route=fn)
Planning	agent.run(prompt, plan=True)
Graph	Graph() with nodes and conditional edges
Retry	retries=3
Timeout	timeout=30.0
Context limit	max_messages=20
Usage tracking	agent.usage

All features are optional. One parameter enables one feature.

Providers

# Mistral (default)
agent = Agent(provider="mistral", model="mistral-large-latest")

# OpenAI
agent = Agent(provider="openai", model="gpt-4o")

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

API keys via environment variables (MISTRAL_API_KEY, OPENAI_API_KEY, ANTHROPIC_API_KEY) or api_key=.

Streaming

async for event in agent.stream("Write a poem"):
    if event.type == "text":
        print(event.content, end="", flush=True)
    elif event.type == "tool_call":
        print(f"\nCalling {event.name}...")

Memory

agent = Agent(session="my-chat")
await agent.run("I'm Jose")
# Later...
await agent.run("What's my name?")  # Remembers

# Manual
agent.save("conversation.json")
agent.load("conversation.json")

Structured outputs

from dataclasses import dataclass

@dataclass
class Analysis:
    sentiment: str
    confidence: float
    keywords: list[str]

result = await agent.run(
    "Analyse this review: 'Great product!'",
    output=Analysis
)

Why

We follow Anthropic's philosophy: start simple, add complexity only when needed. Most agent frameworks do the opposite.

This library gives you a clean starting point that breaks the blank page problem. Instead of figuring out how to structure your agent code, you get a working foundation with sensible defaults. Then you adapt it to your needs.

• ~1,500 lines total. Read the entire codebase in one sitting.
• No hidden behaviour. What you see is what runs.
• Modular by default. Each feature is independent. Use what you need.
• Built to be forked. If it doesn't fit your use case, copy and modify it.

Ideal for

• Projects that need full control. No black boxes.
• Teams that want to understand their code. Not just use it.
• Prototypes that might become products. Start simple, grow as needed.
• Learning how agents work. The code is the documentation.

Contributing

See CONTRIBUTING.md.

Licence

Apache 2.0