Open source · MIT licensed
Coming soon — The framework is under active development. Star the repo to get notified.

The agent framework
that does things

supyagent is an open-source framework for building LLM agents with real-world tools. Define agents in YAML, write tools as Python scripts, and connect to 13+ cloud services with one command.

View on GitHubTry the cloud

Three layers, one ecosystem

Each layer does one thing well. Use them together or independently.

supyagent cloud

The connections

OAuth management, token encryption, auto-refresh, and a unified REST API for 13+ services. Optional — use it when your agent needs third-party access.

supyagent

The brain

The agent conversation loop. YAML-configured agents, any LLM via LiteLLM, session persistence, context management with auto-summarization, and multi-agent delegation.

supypowers

The hands

Tool execution engine. Write Python scripts with Pydantic models — each declares its own dependencies and runs in an isolated uv environment. Zero runtime dependencies.

Tools as Python scripts

Every tool is a standalone Python file. Declare dependencies inline — each tool gets its own isolated uv environment at runtime.

1

No dependency conflicts

A tool using numpy 1.x won't conflict with one using numpy 2.x. Each script resolves independently.

2

Clean agent environment

Your agent's Python environment stays minimal. Tool dependencies never bloat it.

3

Auto-documented

Pydantic models = automatic JSON schemas. Your agent gets typed function signatures with descriptions.

powers/weather.py
# /// script
# dependencies = ["pydantic", "httpx"]
# ///
from pydantic import BaseModel, Field
class WeatherInput(BaseModel):
    city: str = Field(..., description="City name")
class WeatherOutput(BaseModel):
    temperature: float
    conditions: str
def get_weather(input: WeatherInput) -> WeatherOutput:
    """Get current weather for a city."""
    resp = httpx.get(f"https://api.weather.com/{input.city}")
    data = resp.json()
    return WeatherOutput(**data)

Built for real work

Not just a conversation loop. A complete system for building agents that do things.

YAML over code

Define agents declaratively. Model, system prompt, tools, delegates — all in a YAML file. No boilerplate classes to inherit.

YAML
# agents/researcher.yaml
name: researcher
model: gpt-4
system_prompt: |
  You are a research assistant.
  Summarize findings concisely.
powers_path: ./powers
delegates:
  - writer

Any LLM provider

OpenAI, Anthropic, Google, Mistral, Ollama, or any OpenAI-compatible endpoint. Switch models by changing one line in your YAML.

YAML
# Use any provider
model: gpt-4o
model: claude-sonnet-4-5-20250929
model: gemini/gemini-2.0-flash
model: ollama/llama3
model: together_ai/meta-llama/Llama-3-70b

Multi-agent orchestration

Agents delegate to specialists. A manager decomposes tasks and routes to the right agent. Depth limits prevent infinite loops.

YAML
# agents/manager.yaml
name: manager
model: gpt-4
delegates:
  - researcher
  - writer
  - coder
max_delegation_depth: 3

Context management

Full history persists to disk. LLM calls get a trimmed window: system prompt + optional summary + recent messages. Auto-summarization keeps costs down.

YAML
# Built-in context management
context:
  max_messages: 50
  max_tokens: 8000
  auto_summarize: true
  summary_model: gpt-4o-mini

Get started

Two paths to a working agent.

Full-stack app (recommended)

terminal
$ npx create-supyagent-app
$ cd my-supyagent-app
$ npm run dev              # chat UI + agent ready

Python framework

terminal
$ pip install supyagent
$ supyagent init         # scaffold agents/ and powers/
$ supyagent connect      # link cloud integrations
$ supyagent chat myagent  # start chatting
Star on GitHubTry the cloud