240ms microVMs
for the agents you ship.
Open-source Firecracker sandboxes for AI agents and code-interpreters. Full Linux. Persistent volumes. Fork in milliseconds.
The fastest E2B alternative — runs on your cloud or ours.
snapshot restore · 50-boot p50
Powers code execution for
Built for agents that ship to production.
The infrastructure E2B should have been — faster, cheaper, fully open-source, and shape-shifts to your stack.
240ms cold start
Pre-warmed netns pools, XFS reflink CoW rootfs, and snapshot restore. p50 = 240ms · p99 = 251ms.
True KVM isolation
Each sandbox is a real Firecracker microVM with its own kernel. Not a container.
Fork in 64ms
Snapshot any sandbox, fork N copies instantly. Perfect for parallel agent rollouts and search trees.
Persistent volumes
Named volumes survive sandbox death. Hot-swap between flavors. Mount at any path.
Auto-TTL + hibernate
Default 5-min idle. Pass ttl=24h for long runs. Hibernate to disk, wake on demand.
First-class templates
code-interpreter, Next.js, Vite, browser, Claude Code, Codex — or bake your own.
vCPU pinning
Pinned to dedicated cores for stable p99. No noisy-neighbor surprises.
Self-host or managed
Apache-2.0. Run on your own bare metal, GCP, AWS — or use our managed cloud.
A boot path engineered for milliseconds.
No Docker daemon. No image pulls. No init system warm-up. Just KVM and pre-staged state.
Five lines of code from zero to sandboxed.
A simple SDK that abstracts the heavy infra. Python, TypeScript, and a clean REST API.
One SDK, every runtime
Spin up Python, Node, browsers, or Claude Code with the same call. The SDK handles auth, ports, files, streaming, and lifecycle.
sb.fs.*— POSIX-ish FS over HTTPsb.run / execStream— sync + streaming execsb.fork()— instant clone for parallelismsb.expose(port)— public URL with JWTsb.hibernate / wake— zero-cost idle- Events stream over SSE / WebSocket
from pandastack import Sandbox
# 240ms cold-start — no Docker, no waiting
sb = Sandbox.create(template="code-interpreter")
result = sb.run("""
import pandas as pd, numpy as np
df = pd.DataFrame({'x': np.random.randn(1000)})
print(df.describe())
""")
print(result.stdout)
# Snapshot once → fork 50 agents in parallel
forks = [sb.fork() for _ in range(50)]
The fastest sandbox primitive, anywhere.
Measured end-to-end: POST /sandboxes → ssh-ready. us-central1 host, 50-boot p50.
cold start (ms) — lower is better
may 2026# numbers from public docs / first-party measurements.
# mileage varies by region and template size.
Run sandboxes close to your agents.
Eight regions today. Bring-your-own-cloud for everywhere else. Latency-routed by default.
- iadus-east-1Ashburn
- sfous-west-1San Francisco
- ordus-central-1Chicago
- fraeu-central-1Frankfurt
- lhreu-west-2London
- nrtap-northeast-1Tokyo
- sydap-southeast-2Sydney
- grusa-east-1São Paulo
Pick a base. Ship in minutes.
Production-baked snapshots for every common agent stack. Or bring your own Dockerfile.
Code Interpreter
Python 3.12 + pandas, numpy, matplotlib, scipy, jupyter.
OpenAI Agents SDK
Agents SDK + tools pre-installed. Bring your API key.
Custom build
Bring a Dockerfile or build script. We bake the snapshot.
Spin up your first sandbox in 240 milliseconds.
Free tier includes 100 sandbox-hours/month. No credit card. Self-host with Apache-2.0 anytime.
curl -sSL pandastack.ai/install | sh · Linux / macOS · 10MB binary