Instead of spending days wiring together LLMs, tools, and execution environments, Agntrick gives you a production-ready setup instantly.

• Write Less, Do More: Create a fully functional agent with just 5 lines of Python using the zero-config @AgentRegistry.register decorator.
• Context is King (MCP): Native integration with Model Context Protocol (MCP) servers to give your agents live data (Web search, APIs, internal databases).
• Hardcore Local Tools: Built-in blazing fast tools (ripgrep, fd, AST parsing) so your agents can explore and understand local codebases out-of-the-box.
• Stateful & Resilient: Powered by LangGraph to support memory, cyclic reasoning, and human-in-the-loop workflows.
• Docker-First Isolation: Every agent runs in isolated containers—no more "it works on my machine" when sharing with your team.

pip install agntrick

# Or with development dependencies
pip install "agntrick[dev]"

git clone https://github.com/jeancsil/agntrick.git
cd agntrick
make install

You need an LLM API key to breathe life into your agents. Agntrick supports 10+ LLM providers via LangChain!

# Copy the template
cp .env.example .env

# Edit .env and paste your API key
# Choose one of the following providers:
# OPENAI_API_KEY=sk-your-key-here
# ANTHROPIC_API_KEY=sk-ant-your-key-here
# GOOGLE_API_KEY=your-google-key
# GROQ_API_KEY=gsk-your-key-here
# MISTRAL_API_KEY=your-mistral-key-here
# COHERE_API_KEY=your-cohere-key-here

# For Ollama (local), no API key needed:
# OLLAMA_BASE_URL=http://localhost:11434

# List all available agents
agntrick list

# Run an agent with input
agntrick developer -i "Explain this codebase"

# Or try the learning agent with web search
agntrick learning -i "Explain quantum computing in simple terms"

# Anthropic (Recommended)
ANTHROPIC_API_KEY=sk-ant-your-key-here

# OpenAI
OPENAI_API_KEY=sk-your-key-here

# Google GenAI / Vertex
GOOGLE_API_KEY=your-google-key
GOOGLE_VERTEX_PROJECT_ID=your-project-id

# Mistral AI
MISTRAL_API_KEY=your-mistral-key-here

# Cohere
COHERE_API_KEY=your-cohere-key-here

# Azure OpenAI
AZURE_OPENAI_API_KEY=your-azure-key
AZURE_OPENAI_ENDPOINT=https://your-resource.openai.azure.com

# AWS Bedrock
AWS_PROFILE=your-profile

# Ollama (Local, no API key needed)
OLLAMA_BASE_URL=http://localhost:11434

# Hugging Face
HUGGINGFACEHUB_API_TOKEN=your-hf-token

Agntrick includes several pre-built agents for common use cases:

📖 See docs/agents.md for detailed information about each agent.

Fast, zero-dependency tools for working with local codebases:

📖 See docs/tools.md for detailed documentation of each tool.

Model Context Protocol servers for extending agent capabilities:

📖 See docs/mcp-servers.md for details on each server and how to add custom MCP servers.

Agntrick supports 10 LLM providers out of the box, covering 90%+ of the market:

📖 See docs/llm-providers.md for detailed setup instructions.

from agntrick import AgentBase, AgentRegistry

@AgentRegistry.register("my-agent", mcp_servers=["fetch"])
class MyAgent(AgentBase):
    @property
    def system_prompt(self) -> str:
        return "You are my custom agent with the power to fetch websites."

Boom. Run it instantly:

agntrick my-agent -i "Summarize https://example.com"

Want to add your own Python logic? Easy.

from langchain_core.tools import StructuredTool
from agntrick import AgentBase, AgentRegistry

@AgentRegistry.register("data-processor")
class DataProcessorAgent(AgentBase):
    @property
    def system_prompt(self) -> str:
        return "You process data files like a boss."

    def local_tools(self) -> list:
        return [
            StructuredTool.from_function(
                func=self.process_csv,
                name="process_csv",
                description="Process a CSV file path",
            )
        ]

    def process_csv(self, filepath: str) -> str:
        # Magic happens here ✨
        return f"Successfully processed {filepath}!"

Agntrick can be configured via a .agntrick.yaml file in your project root or home directory:

# .agntrick.yaml
llm:
  provider: anthropic  # or openai, google, ollama, etc.
  model: claude-sonnet-4-6  # optional model override
  temperature: 0.7

mcp:
  servers:
    - fetch
    - web-forager

logging:
  level: INFO
  file: logs/agent.log

Command your agents directly from the terminal.

# 📋 List all registered agents
agntrick list

# 🕵️ Get detailed info about what an agent can do
agntrick info developer

# 🚀 Run an agent with input
agntrick developer -i "Analyze the architecture of this project"

# ⏱️ Run with an execution timeout (seconds)
agntrick developer -i "Refactor this module" -t 120

# 📝 Run with debug-level verbosity
agntrick developer -i "Hello" -v

# 📜 View logs
tail -f logs/agent.log

Under the hood, we seamlessly bridge the gap between user intent and execution:

flowchart TB
    subgraph User [👤 User Space]
        Input[User Input]
    end

    subgraph CLI [💻 CLI - agntrick]
        Typer[Typer Interface]
    end

    subgraph Registry [📋 Registry]
        AR[AgentRegistry]
        AD[Auto-discovery]
    end

    subgraph Agents [🤖 Agents]
        Dev[developer agent]
        Learning[learning agent]
        News[news agent]
    end

    subgraph Core [🧠 Core Engine]
        AB[AgentBase]
        LG[LangGraph Runtime]
        CP[(Checkpointing)]
    end

    subgraph Tools [🧰 Tools & Skills]
        LT[Local Tools]
        MCP[MCP Tools]
    end

    subgraph External [🌍 External World]
        LLM[LLM API]
        MCPS[MCP Servers]
    end

    Input --> Typer
    Typer --> AR
    AR --> AD
    AR -->|Routes to| Dev & Learning & News

    Dev & Learning & News -->|Inherits from| AB

    AB --> LG
    LG <--> CP
    AB -->|Uses| LT
    AB -->|Uses| MCP

    LT -->|Reasoning| LLM
    MCP -->|Queries| MCPS
    MCPS -->|Provides Data| LLM

    LLM --> Output[Final Response]

# Install dependencies (blazingly fast with uv ⚡)
make install

# Run the test suite
make test

# Run agents directly
agntrick developer -i "Hello"

make install    # Install dependencies with uv
make test       # Run pytest with coverage
make format     # Auto-format codebase with ruff
make check      # Strict linting (mypy + ruff)
make build      # Build wheel and sdist packages
make build-clean # Remove build artifacts

# Release core agntrick package
make release VERSION=0.3.0

# Release agntrick-whatsapp package
make release-whatsapp VERSION=0.4.0

# Release both packages with different versions
make release-both CORE=0.3.0 WHATSAPP=0.4.0

We love contributions! Check out our AGENTS.md for development guidelines.

For maintainers: See RELEASING.md for how to publish new versions to PyPI.

The Golden Rules:

1. make check should pass without complaints.
2. make test should stay green.
3. Don't drop test coverage (we like our 80% mark!).

This project is licensed under the MIT License. See LICENSE for details.

Stand on the shoulders of giants:

If you find this useful, please consider giving it a ⭐ or buying me a coffee!