minionx 0.1.21

👷‍♂️Minion is Agent's Brain. Minion is designed to execute any type of queries, offering a variety of features that demonstrate its flexibility and intelligence.

Minion

High performance agent framework that can do everything. Minion is designed to execute any type of queries, offering a variety of features that demonstrate its flexibility and intelligence.

Features

• CodeAgent - Python code execution agent with tool calling support
• Tool Search Tool (TST) - Dynamic tool discovery, 85% token reduction (docs)
• Auto-compact - Automatic context window management via history summarization (docs)
• Auto-decay - Large tool response management, saves to files after TTL (docs)
• Multi-provider Support - OpenAI, Azure, Bedrock, Anthropic, LiteLLM (100+ providers)
• Skills - Modular capabilities to extend agent functionality (docs)
• MCP Integration - Model Context Protocol tool support
• Streaming - Real-time streaming responses with structured chunks

Installation

Install from PyPI

pip install minionx

# With optional dependencies
pip install minionx[litellm]      # LiteLLM support (100+ LLM providers)
pip install minionx[anthropic]   # Anthropic Claude
pip install minionx[bedrock]     # AWS Bedrock
pip install minionx[gradio]      # Gradio web UI
pip install minionx[all]         # All optional dependencies

Install from Source

git clone https://github.com/femto/minion.git && cd minion
pip install -e .
cp config/config.yaml.example config/config.yaml
cp config/.env.example config/.env

Configure

Edit config.yaml (see Configuration for file location):

models:
  "default":
    api_type: "openai"
    base_url: "${DEFAULT_BASE_URL}"
    api_key: "${DEFAULT_API_KEY}"
    model: "gpt-5.1"
    temperature: 0

Set your API key:

export DEFAULT_BASE_URL=your-base-url use https://api.openai.com/v1 if openai 
export DEFAULT_API_KEY=your-api-key

See Configuration for more details on configuration options.

Docker Installation

git clone https://github.com/femto/minion.git && cd minion
cp config/config.yaml.example config/config.yaml

# Set your API key
export DEFAULT_BASE_URL=your-base-url use https://api.openai.com/v1 if openai 
export DEFAULT_API_KEY=your-api-key

# Build and run (basic install)
docker-compose build
docker-compose run --rm minion

# Build with optional dependencies
docker-compose build --build-arg EXTRAS="gradio,web,anthropic"
# Or install all extras
docker-compose build --build-arg EXTRAS="all"

# Run a specific example
docker-compose run --rm minion python examples/mcp/mcp_agent_example.py

Quick Start

Using CodeAgent (Recommended)

from minion.agents.code_agent import CodeAgent

# Create agent
agent = await CodeAgent.create(
    name="Minion Code Assistant",
    llm="your-model",
    tools=all_tools,  # optional
)

# Run task
async for event in await agent.run_async("your task here"):
    print(event)

See examples/mcp/mcp_agent_example.py for a complete example with MCP tools.

Using Brain

from minion.main.brain import Brain

brain = Brain()
obs, score, *_ = await brain.step(query="what's the solution 234*568")
print(obs)

See Brain Usage Guide for more examples.

Quick Demo

Click to watch the demo video on YouTube.

Working Principle

The flowchart demonstrates the complete process from query to final result:

1. First receives the user query (Query)
2. System generates a solution (Solution)
3. Performs solution verification (Check)
4. If unsatisfactory, makes improvements (Improve) and returns to generate new solutions
5. If satisfactory, outputs the final result (Final Result)

Documentation

• CodeAgent Documentation - Powerful Python code execution agent
• Brain Usage Guide - Using brain.step() for various tasks
• Skills Guide - Extend agent capabilities with modular skills
• Benchmarks - Performance results on GSM8K, Game of 24, AIME, Humaneval
• Route Parameter Guide - Route options for different reasoning strategies
• Tool Search Tool (TST) - Dynamic tool discovery for large tool libraries (85% token reduction)
• Auto-compact Guide - Automatic context window management via history summarization
• Auto-decay Guide - Automatic context management for large tool responses (Experimental)

Configuration

Configuration File Locations

1. Project Config: MINION_ROOT/config/config.yaml - Default project configuration (see MINION_ROOT)
2. User Config: ~/.minion/config.yaml - User-specific overrides

Configuration Priority

When both configuration files exist:

• Project Config takes precedence over User Config

This allows you to:

• Keep sensitive data (API keys) in your user config
• Share project defaults through the project config

Environment Variables

Variable Substitution: Use ${VAR_NAME} syntax to reference environment variables directly in config values:

models:
  "default":
    api_key: "${OPENAI_API_KEY}"
    base_url: "${OPENAI_BASE_URL}"
    api_type: "openai"
    model: "gpt-4.1"
    temperature: 0.3
  "azure-gpt-4o":
    api_type: "azure"
    api_key: "${AZURE_OPENAI_API_KEY}"
    base_url: "${AZURE_OPENAI_ENDPOINT}"  # e.g., https://your-resource.openai.azure.com/
    api_version: "2024-06-01"
    model: "gpt-4o"  # deployment name
    temperature: 0

Loading .env Files: Use env_file to load environment variables from .env files (follows Docker .env file format):

env_file:
  - .env        # loaded first
  - .env.local  # loaded second, can override values from .env

Inline Environment Variables: Define environment variables directly in config:

environment:
  MY_VAR: "value"
  ANOTHER_VAR: "another_value"

Variables from all sources (system environment, .env files, inline environment) will be available for ${VAR_NAME} substitution throughout the configuration.

Supported API Types

api_type	Description	Required Fields
openai	OpenAI API or compatible (Ollama, vLLM, LocalAI)	api_key, base_url, model
azure	Azure OpenAI Service	api_key, base_url, api_version, model
azure_inference	Azure AI Model Inference (DeepSeek, Phi)	api_key, base_url, model
azure_anthropic	Azure hosted Anthropic models	api_key, base_url, model
bedrock	AWS Bedrock (sync)	access_key_id, secret_access_key, region, model
bedrock_async	AWS Bedrock (async, better performance)	access_key_id, secret_access_key, region, model
litellm	Unified interface for 100+ providers	api_key, model (with provider prefix)

LiteLLM Model Prefixes: Use anthropic/claude-3-5-sonnet, bedrock/anthropic.claude-3, gemini/gemini-1.5-pro, ollama/llama3.2, etc. See LiteLLM docs for all supported providers.

See config/config.yaml.example for complete examples of all supported providers.

Warning: Be cautious - LLM can generate potentially harmful code.

MINION_ROOT

MINION_ROOT is the base directory for locating configuration files (MINION_ROOT/config/config.yaml).

How MINION_ROOT is Determined

1. Checks MINION_ROOT environment variable (if set)
2. Auto-detects by finding .git, .project_root, or .gitignore in parent directories
3. Falls back to current working directory

MINION_ROOT by Installation Method

Installation Method	MINION_ROOT Value
pip install minionx (PyPI)	Your application's current working directory (cwd)
pip install -e . (Source)	The minion source code directory

Example:

# PyPI install - config at /home/user/myproject/config/config.yaml
cd /home/user/myproject
python my_app.py  # MINION_ROOT = /home/user/myproject

# Source install from /home/user/minion
cd /home/user/minion && pip install -e .
cd /anywhere
python my_app.py  # MINION_ROOT = /home/user/minion

Verify MINION_ROOT - Check the startup log:

INFO | minion.const:get_minion_root:44 - MINION_ROOT set to: <some_path>

Related Projects

• minion-agent Production agent system with multi-agent coordination, browser automation, and research capabilities
• minion-code Minion's implementation of Claude Code

Community and Support

WeChat Group (minion-agent discussion):

Optional Dependencies

The project uses optional dependency groups to avoid installing unnecessary packages. Install only what you need:

# Development tools (pytest, black, ruff)
pip install -e ".[dev]"

# LiteLLM - unified interface for 100+ LLM providers
pip install -e ".[litellm]"

# Google ADK and LiteLLM support
pip install -e ".[google]"

# Browser automation (browser-use)
pip install -e ".[browser]"

# Gradio web UI
pip install -e ".[gradio]"

# UTCP support
pip install -e ".[utcp]"

# AWS Bedrock support
pip install -e ".[bedrock]"

# Anthropic Claude support
pip install -e ".[anthropic]"

# Web tools (httpx, beautifulsoup4, etc.)
pip install -e ".[web]"

# Install ALL optional dependencies
pip install -e ".[all]"

# You can also combine multiple groups:
pip install -e ".[dev,gradio,anthropic,litellm]"