SmolAgent

A lightweight AI agent system powered by SmolAgents and Ollama that executes tasks in a sandboxed Docker environment.

Features

• Shell command execution in isolated Docker containers
• File writing and management
• Web search capabilities via DuckDuckGo
• Uses Qwen 2.5 Coder (1.5B) model for fast, efficient operation
• Interactive task-based interface

Requirements

• Python 3.9+
• Docker
• Ollama with qwen2.5-coder:1.5b-instruct model

Installation

1. Install dependencies:

pip install smolagents docker litellm

2. Pull the required model in Ollama:

ollama pull qwen2.5-coder:1.5b-instruct

3. Ensure Docker is running on your system

Usage

Run the agent:

python devin.py

Enter tasks when prompted. The agent will:

• Execute shell commands in the sandbox
• Write files to the workspace directory
• Search the web for information
• Verify results before providing answers

Type exit or quit to stop the agent.

Docker Setup

The agent automatically creates a Docker container named smol_devin_sandbox. The workspace directory is mounted to /app inside the container for file access.

To build a custom image (optional):

docker build -t smol-devin-image .

Configuration

Edit these variables in devin.py to customize:

• MODEL_ID: Change the LLM model
• API_BASE: Ollama server URL
• CONTAINER_NAME: Docker container name
• WORKSPACE_DIR: Local workspace path