The cloud for AI agents.open source, Apache-2.0

A real computer for every agent.

Every job gets its own Firecracker microVM in 50–100 ms: agent sandboxes, Postgres, apps and functions on one substrate. Run it, fork it, throw it away — $0 compute when idle.

agent loop

› fix the failing date parser

agent → tool call: run_code

sandbox created

$ pytest tests/test_dates.py

1 failed · patching parser.py

$ pytest tests/test_dates.py

14 passed

sandbox killed · $0

runs every agent

  • Claude Code
  • Codex
  • Devin
  • OpenHands
  • SWE-agent
  • LangChain
  • CrewAI
  • AutoGen
microVM boot
50–100ms
full Linux kernel · snapshot restore
boot 50–100blink0100200300 ms
same-host fork
~400ms
fork it, snapshot it, throw it away
create 50–100fork ~400 ms0250500 ms
Postgres 16, ready
~34s
a dedicated VM and durable volume
createready0102034 s
compute when idle
$0
per-second metered billing
active: metered per secondidle: $0

Not just a codebox. Four primitives.

Sandboxes, app hosting, Postgres and functions — one Firecracker substrate, four ways to use it.

POST /v1/sandboxes

create 50–100 msfork~400 msfork ~400 mssnapshotkillfork of a fork · runningforkparentthrown awaytimeparallel universes, one host

Sandboxes

Kernel-isolated VMs. Created in 50–100 ms, forked into parallel universes.

create 50–100 ms · fork ~400 ms, same host

POST /v1/apps

gitbuildliveurlgit pushgit pushbuild VMbuild VMhealth checkhealth checkdeploy 1supersededdeploy 2liveone stable URL across deploys

App hosting

git push → built in a VM → live behind a stable URL.

per pull request: spin up · preview · tear down

POST /v1/databases

dedicated microVMpostgres 16pgvectordurable volumecreateready ~34 smainbranchSELECT id FROM docsORDER BY embedding <=> $1LIMIT 5;pgvector included

Postgres 16

A dedicated VM and durable volume per database. pgvector included.

postgres://…db.pandastack.ai · ready in ~34 s

POST /v1/functions

a whole directorymy-fn/main.pyrequirements.txtlib/util.pydeployhttpon demandPOST /v1/functions/{id}/invokecronevery 5 minutes*/5 * * * *next run:00:15:30:45:00between invocations$0 compute

Functions

Deploy whole directories. Invoke over HTTP or on a cron.

HTTP or cron · $0 compute when idle

How fast is 100 ms?

The slow end of a typical boot. Still three of them inside one blink.

one mark = one 100 ms boot · one full row = 2.5 sreal time 30.00 s

One snapshot restore. Eight steps.

No warm pool. Every create walks the same hot path.

16,384

netns slots per host

2MiB

hugepage streaming

9.2Gbps

guest network

Pick a base. Ship in minutes.

Production-baked snapshots for every common agent stack. Or bring your own Dockerfile.

pip install pandastack

from pandastack import Sandbox
 
sb = Sandbox.create(template="code-interpreter")snapshot restore · 50–100 ms
out = sb.exec("echo 'hello from a microVM'")runs inside the microVM
print(out.stdout)
sb.kill()billing stops, per second
stdouthello from a microVM

Ship on the millisecond cloud.

Spin up a full Linux sandbox in the time it takes to blink. Fork it, snapshot it, throw it away. Then do it ten thousand more times.

Free tier with $5.40/mo usage credit. No card. Apache-2.0.

each mark is one eight-step bootreal time: one every 100 ms