Millisecond sandboxes for AI agents.
Podyard is a hosted platform for sandboxes aimed at AI agents. Pods start in about 12 ms on V8 isolates. Snapshot, fork and replay any session.
- cold start on a V8 isolate
- 12 ms
- to fork a running pod
- 3 ms
- warm exec latency
- 4 ms
Install #
One command installs the CLI on macOS or Linux.
curl -fsSL https://podyard.dev/install.sh | shThen set your key with export PODYARD_API_KEY=<your-key>. Windows, npm, pip and Homebrew options are on the install page.
Benchmarks #
Podyard starts a pod in 12 ms. Lower is better.
- Podyard12 ms
- Unikraft10 ms (claim)
- Blaxel25 ms (claim)
- gVisor45 ms (est.)
- Firecracker125 ms (est.)
- E2B150 ms (est.)
- Modal350 ms (est.)
- Fly Machines800 ms (est.)
Podyard: engine benchmark suite on Apple M-series laptop, engine commit 58e3b27, recorded 2026-11-14. Other providers: vendor-reported or estimated, as of 2026-11-14, and not independently verified by Podyard.
Compare #
| provider | model | cold start | fork | snapshot | replay | open source | free tier |
|---|---|---|---|---|---|---|---|
| Podyard | hosted V8 isolates | 12 ms | Yes | Yes | Yes | Engine (Apache-2.0) | Yes |
| E2B | open-source sandbox SDK with hosted sandboxes | 150 ms (est.) | No | Yes | No | Yes | Yes |
| Modal | serverless GPU + CPU functions, container-based | 350 ms (est.) | No | No | No | No | Yes |
| Fly Machines | micro-VMs (Firecracker) on a global edge network | 800 ms (est.) | No | Yes | No | No | No |
| gVisor | application-level sandbox running on the host | 45 ms (est.) | No | No | No | Yes | Yes |
| Firecracker | open-source micro-VM monitor (raw) | 125 ms (est.) | No | Yes | No | Yes | Yes |
| boat.dev | persistent Ubuntu VMs with SSH, Docker and snapshot forks (7-day trial) | not published | not published | Yes | not published | not published | not published |
| Blaxel | microVM sandboxes that resume from standby | 25 ms (claim) | not published | not published | No | not published | not published |
| Unikraft | unikernel-based agent environments | 10 ms (claim) | not published | not published | No | not published | not published |
Competitor rows are compiled from vendor docs and public benchmarks on the date above. Figures marked estimate are not independently verified by Podyard. Blaxel and Unikraft publish their own figures, so they are marked (claim). Blaxel measures resume from standby, not a cold boot. boat.dev does not publish a cold-start time.
API #
Create a pod, run a command, snapshot it, and fork it.
import { PodyardClient } from '@podyard/sdk'
const pyd = new PodyardClient({ apiKey: process.env.PODYARD_API_KEY })
const pod = await pyd.pods.create({ template: 'python:3.12' })
await pod.exec('python -V') // runs in a V8 isolate
await pod.snapshot() // capture state
const fork = await pod.fork() // branch from any pointAbout #
Podyard gives AI agents an isolated place to run code. Each pod is a V8 isolate that starts in milliseconds, so an agent can spin up as many isolated environments as it needs without waiting on a container or VM to boot.
Every run is captured on a Tape, so any pod can be replayed deterministically. Fork a running pod to branch from any point without re-running its setup, and build reusable templates for common environments. Network is default-deny: pods can only reach the hosts you allow.
Podyard is a hosted platform built on an open-source engine licensed under Apache-2.0. The platform adds accounts, pricing, and multi-tenant isolation on top.
Capabilities #
58e3b27.| status | capabilities |
|---|---|
| supported | create |
| supported | exec |
| supported | fs |
| supported | snapshot |
| supported | fork |
| supported | destroy |
| supported | templates |
| supported | usage |
| supported | sleep |
| supported | wake |
| supported | checkpoint |
| supported | undo |
| supported | trace |
| supported | info |
| engine commit | 58e3b27 |
| license | Apache-2.0 |
Coming soon: sandbox run, sandbox shell, Firecracker microVMs (Tier 2), isolate → VM promotion, directory-listing HTTP, per-command status endpoint.