sandboxy 0.0.4

Open-source agent simulation and benchmarking platform

Sandboxy

Open-source framework for developing, testing, and benchmarking AI agents in simulated environments.

What is Sandboxy?

Sandboxy provides a local development environment for building and testing AI agent scenarios. Define scenarios in YAML, run them against any LLM, and evaluate the results.

Use cases:

• Agent Development - Build and iterate on AI agent behaviors locally
• Evaluation & Testing - Run scenarios against models and score their performance
• Dataset Benchmarking - Test models against datasets of cases with parallel execution
• Red-teaming - Test for prompt injection, policy violations, and edge cases

Quick Start

Installation

# Using uv (recommended)
pip install uv
uv pip install sandboxy

# Or with pip
pip install sandboxy

Set up API keys

# Add your API key (OpenRouter gives access to 400+ models)
echo "OPENROUTER_API_KEY=your-key-here" >> .env

Initialize a project

mkdir my-evals && cd my-evals
sandboxy init

This creates:

my-evals/
├── scenarios/     # Your scenario YAML files
├── tools/         # Custom tool definitions
├── agents/        # Agent configurations (optional)
├── datasets/      # Test case datasets
└── runs/          # Output from runs

Run a scenario

# Run with a specific model
sandboxy run scenarios/my_scenario.yml -m openai/gpt-4o

# Compare multiple models
sandboxy run scenarios/my_scenario.yml -m openai/gpt-4o -m anthropic/claude-3.5-sonnet

# Run against a dataset
sandboxy run scenarios/my_scenario.yml --dataset datasets/cases.yml -m openai/gpt-4o

Local development UI

# Start the local dev server with UI
sandboxy open

Opens a browser with a local UI for browsing scenarios, running them, and viewing results.

Writing Scenarios

Scenarios are YAML files that define agent interactions:

id: customer-support
name: "Customer Support Test"
description: "Test how an agent handles a refund request"

system_prompt: |
  You are a customer support agent for TechCo.
  Be helpful but follow company policy.

user_prompt: |
  I want a refund for my purchase. Order #12345.

# Define tools the agent can use
tools:
  - name: lookup_order
    description: "Look up order details"
    params:
      order_id:
        type: string
        required: true
    returns: "Order details for {{order_id}}"

# Evaluation criteria
goals:
  - name: acknowledged_request
    description: "Agent acknowledged the refund request"
    check:
      type: contains
      value: "refund"

  - name: looked_up_order
    description: "Agent used the lookup tool"
    check:
      type: tool_called
      tool: lookup_order

scoring:
  max_score: 100

CLI Reference

# Run scenarios
sandboxy run <file.yml> -m <model>           # Run a scenario
sandboxy run <file.yml> -m <model> --runs 5  # Multiple runs
sandboxy run <file.yml> --dataset <data.yml> # Run against dataset

# Development
sandboxy open                    # Start local UI
sandboxy serve                   # API server only (no browser)
sandboxy init                    # Initialize project structure

# Scaffolding
sandboxy new scenario <name>     # Create scenario template
sandboxy new tool <name>         # Create tool library template

# Information
sandboxy list-models             # List available models
sandboxy list-tools              # List available tool libraries
sandboxy info <file.yml>         # Show scenario details

# MCP Integration
sandboxy mcp inspect <command>   # Inspect MCP server tools
sandboxy mcp list                # List known MCP servers

Models

Sandboxy supports 400+ models via OpenRouter, plus direct provider access:

# OpenRouter models (recommended)
sandboxy run scenario.yml -m openai/gpt-4o
sandboxy run scenario.yml -m anthropic/claude-3.5-sonnet
sandboxy run scenario.yml -m google/gemini-pro
sandboxy run scenario.yml -m meta-llama/llama-3-70b

# List available models
sandboxy list-models
sandboxy list-models --search claude
sandboxy list-models --free

MLflow Integration

Export scenario run results to MLflow for experiment tracking and model comparison.

# Install with MLflow support
pip install sandboxy[mlflow]

# Export run to MLflow
sandboxy scenario scenarios/test.yml -m openai/gpt-4o --mlflow-export

# Custom experiment name
sandboxy scenario scenarios/test.yml -m gpt-4o --mlflow-export --mlflow-experiment "my-evals"

Or enable in scenario YAML:

id: my-scenario
name: "My Test"

mlflow:
  enabled: true
  experiment: "agent-evals"
  tags:
    team: "support"

system_prompt: |
  ...

See MLFLOW_TRACKING_URI env variable to configure the MLflow server.

Configuration

Environment variables (in ~/.sandboxy/.env or project .env):

Variable	Description
OPENROUTER_API_KEY	OpenRouter API key (400+ models)
OPENAI_API_KEY	Direct OpenAI access
ANTHROPIC_API_KEY	Direct Anthropic access
MLFLOW_TRACKING_URI	MLflow tracking server URI

Project Structure

sandboxy/
├── sandboxy/           # Python package
│   ├── core/           # Runner, state management
│   ├── scenarios/      # Unified scenario runner
│   ├── datasets/       # Dataset benchmarking
│   ├── agents/         # Agent loading and execution
│   ├── tools/          # Tool loading (YAML tools)
│   ├── providers/      # LLM provider integrations
│   ├── api/            # Local dev API server
│   ├── cli/            # Command-line interface
│   ├── local/          # Local project context
│   └── mcp/            # MCP client integration
└── local-ui/           # Local development UI (React)

Contributing

Contributions welcome! See CONTRIBUTING.md.

License

Apache 2.0 - see LICENSE.