OpenAI Agents SDK FastAPI Starter with Vercel Sandbox

Minimal FastAPI app that runs the OpenAI Agents SDK with Vercel Sandbox on Vercel's Python runtime. Each request spins up an isolated microVM, gives the agent shell access to analyze data, and tears it down when done.

Prerequisites

• Python 3.12+ (Vercel default is 3.12).
• A Vercel account and Vercel CLI (npm i -g vercel).
• An OpenAI API key.

Environment variables

Setup

1
vercel link
2
vercel env pull

This creates .env.local with your VERCEL_OIDC_TOKEN. Add your OpenAI key:

1
echo 'OPENAI_API_KEY=sk-...' >> .env.local

Local development

1
uv sync
2
uv run uvicorn app:app --reload --host 127.0.0.1 --port 8000

Open http://127.0.0.1:8000 to use the interactive demo. The agent has shell access to a sandbox with sample sales data (sales.csv).

API endpoints:

• GET /api/health returns {"status": "ok", "openai_configured": true}
• POST /api/run with {"input": "Which region grew the most?"} runs the sandbox agent

Deploy to Vercel

1
vercel

Vercel detects app.py and the app ASGI instance. Dependencies come from pyproject.toml. Make sure OPENAI_API_KEY is set under Project > Settings > Environment Variables. The VERCEL_OIDC_TOKEN is auto-provisioned on Vercel deployments.

Rate limiting

The demo includes in-memory rate limiting (5 requests/minute per IP). This resets on cold starts and does not persist across function instances.

Timeouts

Sandbox creation and agent runs can take several seconds. Heavy workloads may need Fluid Compute or Vercel Workflow for durable steps.

How it works

1. Each POST /api/run creates a fresh Vercel Sandbox microVM with sample data.
2. A SandboxAgent with Shell capability receives the user's prompt.
3. The agent writes and runs shell commands inside the sandbox to answer the question.
4. The sandbox is torn down after the response is returned.

License

MIT (match your org's policy when publishing).