Try to buildwith Ailoy

Agent SDK, built around a sandbox.
Open source, for Python, Node.js and Rust.

Read the docs

Trusted by

isolation: microVM; Its own kernel, not a container.
image: Any OCI; Or build one with Recipe steps.
gpu: Vulkan; Passed through into the VM.
host: Linux · macOS · Windows; Windows natively, no WSL.
daemon: None; No Docker, no background service.

01 · Run

Ask once.It works on its own machine.

A real run of the SAM3 example.

/examples/sam3 →

Image editor with SAM3done

promptRemove the background of this GIFbaby-turns-around.gif · 99 frames

Read the SAM3 skill /skills/sam3
Opened the GIF 508×640 · 99 frames
Tracked “baby” on the GPU 97/97 frames · 406.6 s
export.py: FileNotFoundError …/obj1/00097.png
Patched it: a frame with no mask gets an empty one
Ran export.py again

Background removed; animation preserved.

baby-no-background.gif3.1 MBbaby-no-background.webp5.2 MB

inside its VM · /artifacts/baby-no-background.webp

waiting for frames…

More examples →

02 · How it works

The agent runs on your computer.Its work runs in a microVM.

Only what it writes to /artifacts comes back.

/docs →

console = await (    ConsoleClient.builder()    .image(        Recipe("python:3.12-slim-trixie")        .step("apt-get install mesa-vulkan-drivers libvulkan1")        .step("pip install ncnn")    )    .mount_readonly("data/ncnn", "/models")    .mount_readonly(        HostMount(Directory().with_file("SKILL.md", …), …),        "/skills/sam3",    )    .mount("artifacts", "/artifacts")    .gpu(True)    .build()) agent = await (    ailoy.AgentBuilder("openai/gpt-6-astra")    .system_tools()    .console(console)    .skill("/skills/sam3")    .build()) async with console:    async for output in agent.run("Remove the background of this GIF"):        print(output["message"])

examples/sam3/python/main.py →

const consoleClient = await ConsoleClient.builder()  .image(    new Recipe('python:3.12-slim-trixie')      .step('apt-get install mesa-vulkan-drivers libvulkan1')      .step('pip install ncnn'),  )  .mountReadonly('data/ncnn', '/models')  .mountReadonly(    new HostMount(new Directory().withFile('SKILL.md', …), …),    '/skills/sam3',  )  .mount('artifacts', '/artifacts')  .gpu(true)  .build() const agent = await new AgentBuilder('openai/gpt-6-astra')  .systemTools()  .console(consoleClient)  .skill('/skills/sam3')  .build() for await (const { message } of agent.run('Remove the background of this GIF')) {  console.log(message)}

examples/sam3/node/main.mjs →

let mut agent = AgentBuilder::new("openai/gpt-6-astra")    .system_tools()    .console(        ConsoleClient::builder()            .image(                Recipe::new("python:3.12-slim-trixie")                    .step("apt-get install mesa-vulkan-drivers libvulkan1")                    .step("pip install ncnn"),            )            .mount_readonly(project.join("data/ncnn"), "/models")            .mount_readonly(                HostMount::try_new(                    Directory::new().with_file("SKILL.md", …)?,                    …,                )?,                "/skills/sam3",            )            .mount(project.join("artifacts"), "/artifacts")            .gpu(true)            .build()            .await?,    )    .skill("/skills/sam3")    .build()    .await?; let mut stream = agent.run(query);while let Some(output) = stream.next().await {    println!("{:?}", output?.message);}

examples/sam3/rust/main.rs →

The microVM comes from virtx, our open-source runtime: disposable Linux VMs created from code, with no Docker or daemon.

03 · Why a VM

Five ways to run an agent’s work,and where each draws the line.

Compared by category, from each category’s public docs. Where a whole category cannot be pinned down, the cell says so.

	Host process	OS sandbox	Container	Hosted sandbox API	Ailoy
e.g.	an SDK’s local runner	Seatbelt- or Landlock-based tools	Docker, Podman	E2B, Daytona, Modal
Where the work runs	your computer	your computer	your computer	the provider’s cloud	your computer
Boundary	none, or file rules only	policy on the host kernel	namespaces on the shared host kernel	varies by provider	its own Linux kernel (microVM)
GPU inside the boundary	no boundary	varies	possible, with setup	varies	Vulkan, passed through
Host OS	any	platform-specific	Linux; a VM on macOS and Windows	any (it is remote)	Linux · macOS · Windows
Background service	none	none	a container daemon	an API and an account	none
Fine-grained policy	none	policy language, deny by default	varies	varies	mounts, network on or off

Host processan SDK’s local runner

OS sandboxSeatbelt- or Landlock-based tools

ContainerDocker, Podman

Hosted sandbox APIE2B, Daytona, Modal

Where the work runs

Ailoyyour computer

Host processyour computer

OS sandboxyour computer

Containeryour computer

Hosted sandbox APIthe provider’s cloud

Boundary

Ailoyits own Linux kernel (microVM)

Host processnone, or file rules only

OS sandboxpolicy on the host kernel

Containernamespaces on the shared host kernel

Hosted sandbox APIvaries by provider

GPU inside the boundary

AiloyVulkan, passed through

Host processno boundary

OS sandboxvaries

Containerpossible, with setup

Hosted sandbox APIvaries

Host OS

AiloyLinux · macOS · Windows

Host processany

OS sandboxplatform-specific

ContainerLinux; a VM on macOS and Windows

Hosted sandbox APIany (it is remote)

Background service

Ailoynone

Host processnone

OS sandboxnone

Containera container daemon

Hosted sandbox APIan API and an account

Fine-grained policy

Ailoymounts, network on or off

Host processnone

OS sandboxpolicy language, deny by default↑

Containervaries

Hosted sandbox APIvaries

Read 2026-10-20.↑ another category is ahead here

04 · Roadmap

Local today.The same machine goes further later.

Each step keeps the agent and its machine as one thing you ship.

1 now
Local runtime
Your agent runs in its own microVM on your machine.
2 planned
Self-hosted fleet
The same machine runs on your servers, many at once.
3 planned
Managed cloud
Every agent gets a computer on demand, run by BrekkyLab.

05 · Start