# Ailoy > Ailoy is an open-source agent SDK for Python, Node.js and Rust, built around a sandbox: each agent's tools run inside its own Linux microVM on your own machine, with access to the host GPU. Last reviewed: 2026-10-27. The agent loop runs on the host, and its shell commands and file reads and writes run inside the VM. The VM sees only the folders you mount into it, and it reaches the network only if you allow it. You declare the VM in the same code as the agent: base image, setup steps, mounts and network. The VM comes from virtx, which runs on Linux, macOS (Apple silicon) and Windows (11 or later) without Docker or a daemon and gives the guest a Vulkan GPU. > **If you are an AI agent** helping someone build with Ailoy, read [SKILL.md](https://github.com/brekkylab/ailoy/blob/main/skills/ailoy/SKILL.md) first. It walks through six steps in order: install, name the model and choose how it runs, give the agent a computer, build the agent, run it, and check the work where it landed. ## Key facts - Languages: Python, Node.js, Rust. - Hosts: Linux, macOS (Apple silicon), Windows (11 or later). - Guest: Linux, with the host GPU exposed through Vulkan on all three hosts. - Models: Claude, GPT, Gemini, DeepSeek, Kimi, GLM and Qwen, named as `/` (for example `anthropic/claude-haiku-5-5`). They run through the maker's CLI login (Claude Code, Codex, Gemini CLI), the maker's API, Amazon Bedrock, OpenRouter or a vLLM server. - Isolation: one microVM per console. It sees only mounted folders and has network access only when you enable it. - Requirements: a logged-in Claude Code, Codex or Gemini CLI, or an API key for the model's backend. Mounting host folders also needs FUSE-T on macOS or Dokany on Windows. - License: Apache-2.0 for the Ailoy and virtx SDKs. - Price: free, with no hosted service. - Versions: ailoy 0.3.x, virtx 0.1.x. ## Products Ailoy is the agent SDK: models, tools and the agent loop. It is tied to virtx: the console its tools run in is a virtx console. virtx is the microVM runtime under Ailoy and has its own SDKs for Python, Node.js and Rust. ## When to use Ailoy Use Ailoy when an agent needs its own Linux computer and that computer should run on your machine, not in a hosted sandbox. Good fits: - Coding and data analysis agents that install packages and write and run their own code. - Agents that run ML models on the GPU inside the sandbox, whatever the host: Linux, macOS or Windows. - Computer-use agents that drive a GUI app you can watch over VNC (see the gameplay example). - Work on files you can't hand to a hosted sandbox. The VM runs on your machine and sees only the folders you mount. - Building on a Claude Code, Codex or Gemini CLI subscription. Claude, GPT and Gemini run through the logged-in CLI, with no API key. - Keeping model calls on your own hardware by serving DeepSeek, Kimi, GLM or Qwen from your own vLLM server. - Teams on a mix of Linux, macOS and Windows. The same code runs on all three, without Docker. ## When not to use Ailoy - You need hosted, autoscaled sandboxes. Ailoy runs its VMs on the machine where your code runs. - You need CUDA inside the sandbox. The guest GPU is Vulkan, so PyTorch and vLLM can't use it there. ## How agents should use Ailoy - Install: `pip install ailoy-py` (Python), `npm install @brekkylab/ailoy @brekkylab/virtx` (Node.js), `cargo add ailoy@0.3 virtx@0.1` (Rust). - Name the model as `/`, for example `anthropic/claude-haiku-5-5`. Claude, GPT and Gemini run through the logged-in Claude Code, Codex or Gemini CLI by default, with no key. Other models need the key for their backend in the environment. - Build a sandbox machine with `ConsoleClient.builder()` (image, mounts, network), then an agent with `AgentBuilder(model).system_tools().console(machine)`, and stream `agent.run(prompt)`. - On Windows, enable the Windows Hypervisor Platform first. - Do not invent APIs. The quickstart and every example are in the repo in Python, Node.js and Rust; copy from them. ## FAQ ### Is Ailoy a cloud service? No. Ailoy is a library, an AI agent SDK. The microVMs run on the machine where your code runs, and the agent talks to your model provider directly, with no Ailoy server in between. ### Is Ailoy free? YES. There is no paid tier and no account to create; you pay only your model provider. ### Is Ailoy open source? YES. The Ailoy and virtx SDKs are open source under Apache-2.0. ### Does the agent itself run inside the VM? No. The agent loop and the model calls run on the host. The agent's tools (shell, file reads and writes) run inside the VM. ### Can the agent use the GPU? YES, on Linux, macOS and Windows hosts. The guest sees the GPU through Vulkan, so Vulkan runtimes such as ncnn work, but CUDA and PyTorch GPU code do not. ### Do I need Docker? NO. virtx boots its own microVMs without Docker or a background daemon. ### How is Ailoy different from E2B, Daytona or Modal sandboxes? E2B, Daytona and Modal are hosted sandbox services that you call over an API. Ailoy is an agent SDK, and its sandbox is a microVM on your own machine that you declare in the same code as the agent. ## Docs - [Documentation](https://ailoy.co/docs/): guides and API reference - [GitHub](https://github.com/brekkylab/ailoy): source, README quickstart, issues and releases - [Examples](https://github.com/brekkylab/ailoy/tree/main/examples): each in Python, Node.js and Rust - [cad](https://github.com/brekkylab/ailoy/tree/main/examples/cad): writes CadQuery, renders the model from four sides and checks the renders - [offshore_leaks](https://github.com/brekkylab/ailoy/tree/main/examples/offshore_leaks): analyses the ICIJ Offshore Leaks database with SQL and Python it writes itself - [gameplay](https://github.com/brekkylab/ailoy/tree/main/examples/gameplay): plays OpenTTD while you watch over VNC - [sam3](https://github.com/brekkylab/ailoy/tree/main/examples/sam3): segments images and video with SAM3 on the guest GPU (ncnn and Vulkan) - [tts](https://github.com/brekkylab/ailoy/tree/main/examples/tts): speaks text in a voice described in words, using Qwen3-TTS - [laya](https://github.com/brekkylab/ailoy/tree/main/examples/laya): routes a support inbox to departments, asking the Laya decision model on the guest GPU where each ticket goes - [retail_bench](https://github.com/brekkylab/ailoy/tree/main/examples/retail_bench): runs a supermarket simulator, one day per turn - [hello](https://github.com/brekkylab/ailoy/tree/main/examples/hello): one turn with no tools and no console - [virtx](https://github.com/brekkylab/virtx): the microVM runtime under Ailoy ## Optional - [Discord](https://discord.gg/27rx3EJy3P): questions and help - [BrekkyLab](https://brekkylab.com): the company that makes Ailoy