One Interface to Run AI Code Across Daytona, E2B, and Docker PolySandbox is a unified sandbox orchestrator that lets you safely run, evaluate, and compare Python code across multiple execution backends — all through one consistent interface. It’s designed for benchmarking AI-generated code from datasets like MBPP and HumanEval, enabling reproducible and backend-agnostic evaluation.

Unified API — One /run endpoint for Daytona, E2B, and Docker

Dataset Integration — Load MBPP and HumanEval tasks for evaluation

FastAPI Backend — Clean async API for sandbox orchestration

Streamlit UI — Simple interface to run and compare results visually

Scoring & Metrics — View stdout, stderr, runtime, and correctness

Extensible Design — Add new sandboxes or datasets easily via adapters

User/UI → FastAPI Server → Evaluator → Sandbox Adapter → Daytona/E2B/Docker

Adapters: Implement a shared SandboxClient interface Evaluator: Normalizes results into an ExecutionResult schema RunnerAgent: Chooses backend and coordinates runs dynamically

1️⃣ Install UV and dependencies uv venv source .venv/bin/activate uv sync

2️⃣ Set up environment variables

Create a .env file (not committed):

DAYTONA_API_KEY=your_key_here E2B_API_KEY=your_key_here

3️⃣ Run the API uv run uvicorn poly_sandbox.main:app --reload

4️⃣ Run the Streamlit UI uv run streamlit run poly_sandbox/ui/app.py

Run all tests:

uv run pytest -v

Run a specific test:

uv run pytest poly_sandbox/tests/test_adapters.py

🧩 Example API Call curl -X POST "http://localhost:8000/run"
-H "Content-Type: application/json"
-d '{"backend":"daytona","code":"print(2+3)","tests":"assert 2+3==5"}'

Response:

{ "stdout": "5", "stderr": "", "success": true, "runtime_ms": 423, "backend": "daytona" }

poly_sandbox/ ├── adapters/ # Daytona, E2B, Docker clients ├── datasets/ # MBPP, HumanEval loaders ├── evaluators/ # Executor + Scorer logic ├── ui/ # Streamlit frontend ├── utils/ # Config, logging ├── tests/ # Pytest suite └── main.py # FastAPI entrypoint

Evaluating AI code safely across sandboxes is fragmented. PolySandbox unifies it — one API, multiple backends, consistent results.

Unified interface for three sandbox systems

End-to-end demo (UI → API → Sandbox → Scorer)

Modular and extensible adapter architecture

Built in under 10 hours for a hackathon demo