Self-hosted · Firecracker microVMs
Sandboxes for agents,
in your own cloud
MicroVMs that hold an agent's memory and files between sessions, running on infrastructure you control.
Requests finish in seconds. Agents work for hours, and pick up where they left off.
A fresh container per task throws that work away. An agent needs a machine it can set down, come back to, and copy when it wants to try something else.
Pause
Freeze memory and files while the agent waits.
Resume
Back where it stopped, about 200 ms later.
Fork
Copy a running sandbox to try another path.
One machine per agent, shared with no one
Build the image once, keep the state, and choose exactly what each sandbox can reach.
Compute
Sessions up to a day, paused as long as you like
Storage
Room to build, state that persists
Networking
Nothing exposed you didn't choose
Your agent inside, your code outside
Install any agent CLI or framework in a template, then drive the machine from Python, TypeScript, or plain HTTP.
Agents that run inside
SDKs and APIs
from e2b import Sandbox
sbx = Sandbox.create("base")
out = sbx.commands.run("python3 -V")
print(out.stdout)
sbx.pause() # memory + files kept
sbx = Sandbox.connect(sbx.sandbox_id)
[twin] = sbx.fork() # same state, new VMOpen-source SDKs · Apache-2.0
You hold the keys, the network, and the data
Open-source parts, a host with nothing listening, and an install that lives in your own account.
Open source
Built on open code
The sandbox runtime and the SDKs are Apache-2.0, so you can read exactly what runs your agents.
Network
No open ports
Traffic arrives over HTTPS through an outbound-only tunnel. Each sandbox needs its own access token.
Your cloud
Deploy in your account
One Terraform apply on Google Cloud, or Docker Compose on any Linux host with KVM.