Skip to main content
Every snippet on these pages is drop-in runnable. Set PODFLARE_API_KEY once (create one on the dashboard) and each page has working code you can paste straight into your editor.

​
What people actually build

Podflare sandboxes are hardware-isolated microVMs. That means each one is a full Linux box — you can install packages, spin up servers, mount filesystems, run ML inference, crawl websites, provision databases. In practice, five workflows come up over and over:

Code-interpreter agents

Claude’s code_execution tool, OpenAI Agents SDK’s function tools, Vercel AI SDK’s tool(). One helper per framework, the same isolated microVM underneath.

Data & analytics

pandas / matplotlib chart-in-reply, DuckDB against remote CSVs and parquet. No DB server, no infra per tenant.

Database work

Ephemeral Postgres per sandbox for testing generated SQL, RLS policies, migrations. Local Supabase (Postgres + PostgREST + Auth) for full-stack app testing.

Web automation

Headless Chromium via Playwright — scrape, screenshot, fill forms. Boot a Next.js dev server and screenshot agent-generated UIs.

Dev environments

Clone a branch, run tests, return results. Perfect drop-in for PR-review bots and “run my test suite” agents.

​
Why a sandbox instead of running agent code locally

Four reasons that show up the moment you have real users:
  1. Isolation. The agent is going to execute code that came from an LLM. You don’t trust the LLM with your production environment, your files, your credentials, or your network.
  2. Per-session state. Each user conversation gets its own filesystem, installed packages, and process tree. No “did user A’s pip install break user B’s session?” debugging.
  3. Horizontal reset. When a session ends, the sandbox is gone — all state, all side effects. No cleanup scripts, no orphaned processes, no accumulated cruft.
  4. Speed. Warm-pool sandbox restore is ~12 ms on bare metal. The sandbox exists before your agent asks for one.

​
Pattern library

These are the small, composable patterns you’ll see in the examples. Most real agents are combinations of two or three of these:
  • One-shot exec — run a single snippet, return stdout/stderr/exit_code.
  • Persistent REPL — keep the sandbox alive across agent turns; DataFrames and files survive.
  • Fork-and-explore — Sandbox.fork() spawns N copies of a running state so the agent can try N paths in parallel, then commit the winner back to the parent.
  • Upload / download — push input files in, pull output files out.
  • Background services — boot Postgres / Next.js / an HTTP API inside the sandbox; tear down with the sandbox.

​
Shared prerequisites

Every example on this site assumes: 
# Python
pip install podflare

# TypeScript
npm install podflare

# The SDK reads PODFLARE_API_KEY automatically.
export PODFLARE_API_KEY=pk_live_...
Mint a key on dashboard.podflare.ai/keys. Free tier gets a $200 credit — more than enough to run every example on this page end-to-end.