TinyAgent

A small, modular agent framework for building LLM-powered applications in Python.

Inspired by smolagents and Pi — borrowing the minimal-abstraction philosophy from the former and the conversational agent loop from the latter.

Beta — TinyAgent is usable but not production-ready. APIs may change between minor versions.

Note: Reference copy of alchemy-rs available at /home/tuna/alchemy-rs-ref

Overview

TinyAgent provides a lightweight foundation for creating conversational AI agents with tool use capabilities. It features:

• Streaming-first architecture: All LLM interactions support streaming responses
• Tool execution: Define and execute tools with structured outputs
• Event-driven: Subscribe to agent events for real-time UI updates
• Provider agnostic: Works with any OpenAI-compatible /chat/completions endpoint (OpenRouter, OpenAI, Chutes, local servers)
• Prompt caching: Reduce token costs and latency with Anthropic-style cache breakpoints
• Provider path: Default built-in Rust (alchemy_provider) with optional proxy integration
• Type-safe: Full type hints throughout

Quick Start

import asyncio
from tinyagent import Agent, AgentOptions
from tinyagent.alchemy_provider import OpenAICompatModel, stream_alchemy_openai_completions

# Create an agent
agent = Agent(
    AgentOptions(
        stream_fn=stream_alchemy_openai_completions,
        session_id="my-session"
    )
)

# Configure
agent.set_system_prompt("You are a helpful assistant.")
agent.set_model(
    OpenAICompatModel(
        provider="openrouter",
        id="anthropic/claude-3.5-sonnet",
        base_url="https://openrouter.ai/api/v1/chat/completions",
    )
)
# Optional: any OpenAI-compatible /chat/completions endpoint
# agent.set_model(OpenAICompatModel(provider="openai", id="gpt-4o-mini", base_url="https://api.openai.com/v1/chat/completions"))

# Simple prompt
async def main():
    response = await agent.prompt_text("What is the capital of France?")
    print(response)

asyncio.run(main())

Installation

pip install tiny-agent-os

Core Concepts

Agent

The Agent class is the main entry point. It manages:

• Conversation state (messages, tools, system prompt)
• Streaming responses
• Tool execution
• Event subscription

Messages

Messages are Pydantic models (use attribute access):

• UserMessage: Input from the user
• AssistantMessage: Response from the LLM
• ToolResultMessage: Result from tool execution

Tools

Tools are functions the LLM can call:

from tinyagent import AgentTool, AgentToolResult, TextContent

async def calculate_sum(tool_call_id: str, args: dict, signal, on_update) -> AgentToolResult:
    result = args["a"] + args["b"]
    return AgentToolResult(
        content=[TextContent(text=str(result))]
    )

tool = AgentTool(
    name="sum",
    description="Add two numbers",
    parameters={
        "type": "object",
        "properties": {
            "a": {"type": "number"},
            "b": {"type": "number"}
        },
        "required": ["a", "b"]
    },
    execute=calculate_sum
)

agent.set_tools([tool])

Events

The agent emits events during execution:

• AgentStartEvent / AgentEndEvent: Agent run lifecycle
• TurnStartEvent / TurnEndEvent: Single turn lifecycle
• MessageStartEvent / MessageUpdateEvent / MessageEndEvent: Message streaming
• ToolExecutionStartEvent / ToolExecutionUpdateEvent / ToolExecutionEndEvent: Tool execution

Subscribe to events:

def on_event(event):
    print(f"Event: {event.type}")

unsubscribe = agent.subscribe(on_event)

Prompt Caching

TinyAgent supports Anthropic-style prompt caching to reduce costs on multi-turn conversations. Enable it when creating the agent:

agent = Agent(
    AgentOptions(
        stream_fn=stream_alchemy_openai_completions,
        session_id="my-session",
        enable_prompt_caching=True,
    )
)

Cache breakpoints are automatically placed on user message content blocks so the prompt prefix stays cached across turns. See Prompt Caching for details.

Rust Binding: tinyagent._alchemy

TinyAgent ships with an optional Rust-based LLM provider implemented in src/lib.rs. It wraps the alchemy-llm Rust crate and exposes it to Python via PyO3 as tinyagent._alchemy, giving you native-speed OpenAI-compatible streaming without leaving the Python process.

Why

The pure-Python proxy path (proxy.py) remains available, while the Rust path provides:

• Lower per-token overhead -- SSE parsing, JSON deserialization, and event dispatch all happen in compiled Rust with a multi-threaded Tokio runtime.
• Unified provider abstraction -- alchemy-llm normalizes differences across providers (OpenRouter, Anthropic, custom endpoints) behind a single streaming interface.
• Full event fidelity -- text deltas, thinking deltas, tool call deltas, and terminal events are surfaced as TinyAgent event/message models.

How it works

Python (async)             Rust (Tokio)
─────────────────          ─────────────────────────
stream_alchemy_*()  ──>    alchemy_llm::stream()
                            │
AlchemyStreamResponse       ├─ SSE parse + deserialize
  .__anext__()       <──    ├─ event_to_py_value()
  (asyncio.to_thread)       └─ mpsc channel -> Python

1. Python calls openai_completions_stream(model, context, options) which is a #[pyfunction].
2. The Rust side builds an alchemy-llm request, opens an SSE stream on a shared Tokio runtime, and sends events through an mpsc channel.
3. Python reads events by calling the blocking next_event() method via asyncio.to_thread, making it async-compatible without busy-waiting.
4. A terminal done or error event signals the end of the stream. The final AssistantMessage dict is available via result().

Building

Requires a Rust toolchain (1.70+) and maturin.

pip install maturin
maturin develop            # debug build, installs into current venv
maturin develop --release  # optimized build

Python API

Two functions are exposed from the tinyagent._alchemy module:

Function	Description
collect_openai_completions(model, context, options?)	Blocking. Consumes the entire stream and returns {"events": [...], "final_message": {...}}. Useful for one-shot calls.
openai_completions_stream(model, context, options?)	Returns an OpenAICompletionsStream handle for incremental consumption.

The OpenAICompletionsStream handle has two methods:

Method	Description
next_event()	Blocking. Returns the next event dict, or None when the stream ends.
result()	Blocking. Returns the final assistant message dict.

All three arguments are plain Python dicts:

model = {
    "id": "anthropic/claude-3.5-sonnet",
    "base_url": "https://openrouter.ai/api/v1/chat/completions",
    "provider": "openrouter",          # required for env-key fallback/inference
    "api": "openai-completions",       # optional; inferred from provider when omitted/blank
    "headers": {"X-Custom": "val"},   # optional
    "reasoning": False,                  # optional
    "context_window": 128000,            # optional
    "max_tokens": 4096,                  # optional
}

context = {
    "system_prompt": "You are helpful.",
    "messages": [
        {"role": "user", "content": [{"type": "text", "text": "Hello"}]}
    ],
    "tools": [                       # optional
        {"name": "sum", "description": "Add numbers", "parameters": {...}}
    ],
}

options = {
    "api_key": "sk-...",            # optional
    "temperature": 0.7,              # optional
    "max_tokens": 1024,              # optional
}

Routing contract (provider, api, base_url):

• provider: backend identity used for API-key fallback and provider defaults
• api: alchemy unified API selector (openai-completions or minimax-completions)
• base_url: concrete HTTP endpoint

If api is omitted/blank, the Python side infers:

• provider in {"minimax", "minimax-cn"} => minimax-completions
• otherwise => openai-completions

Legacy API aliases are normalized for backward compatibility:

• api="openrouter" / api="openai" => openai-completions
• api="minimax" => minimax-completions

Using via TinyAgent (high-level)

You don't need to call the Rust binding directly. Use the alchemy_provider module:

from tinyagent import Agent, AgentOptions
from tinyagent.alchemy_provider import OpenAICompatModel, stream_alchemy_openai_completions

agent = Agent(
    AgentOptions(
        stream_fn=stream_alchemy_openai_completions,
        session_id="my-session",
    )
)
agent.set_model(
    OpenAICompatModel(
        provider="openrouter",
        id="anthropic/claude-3.5-sonnet",
        base_url="https://openrouter.ai/api/v1/chat/completions",
    )
)

MiniMax global:

agent.set_model(
    OpenAICompatModel(
        provider="minimax",
        id="MiniMax-M2.5",
        base_url="https://api.minimax.io/v1/chat/completions",
        # api is optional here; inferred as "minimax-completions"
    )
)

MiniMax CN:

agent.set_model(
    OpenAICompatModel(
        provider="minimax-cn",
        id="MiniMax-M2.5",
        base_url="https://api.minimax.chat/v1/chat/completions",
        # api is optional here; inferred as "minimax-completions"
    )
)

Cross-provider tool-call smoke examples:

• One-agent workflow: examples/example_tool_calls_three_providers.py
• Raw Rust binding workflow (multi-turn tools): scripts/smoke_rust_tool_calls_three_providers.py
  • Command: uv run python scripts/smoke_rust_tool_calls_three_providers.py

Limitations

• Rust binding currently dispatches only openai-completions and minimax-completions.
• Image blocks are not yet supported (text and thinking blocks work).
• next_event() is blocking and runs in a thread via asyncio.to_thread -- this adds slight overhead compared to a native async generator, but keeps the GIL released during the Rust work.

Documentation

• Architecture: System design and component interactions
• API Reference: Detailed module documentation
• Prompt Caching: Cache breakpoints, cost savings, and provider requirements
• OpenAI-Compatible Endpoints: Using OpenAICompatModel.base_url with OpenRouter/OpenAI/Chutes-compatible backends
• Usage Semantics: Canonical usage schema across provider flows
• Changelog: Release history

Project Structure

tinyagent/
├── agent.py              # Agent class
├── agent_loop.py         # Core agent execution loop
├── agent_tool_execution.py  # Tool execution helpers
├── agent_types.py        # Type definitions
├── caching.py            # Prompt caching utilities
├── alchemy_provider.py   # Rust-based provider (PyO3)
├── proxy.py              # Proxy server integration
└── proxy_event_handlers.py  # Proxy event parsing