Ailoy

Ailoy is a library for building any kind of AI agents right in your own code.

It supports Python, JavaScript ( TypeScript), and Rust APIs, and works on Linux, Windows, and macOS.

[!WARNING] Ailoy is under active development, and its API may change between versions.

Requirements

No system dependencies are required, but your agent's model needs a way to run: by default its provider's CLI (Claude Code, Codex or Gemini CLI), or its API with a key.

Mounting host folders into the console (virtx's virtual filesystem) needs FUSE: install FUSE-T on macOS or Dokany on Windows. See the virtx documentation for details.

On Windows, the console's micro-VM runs on the Windows Hypervisor Platform, which is off by default. Turn the optional feature on from an administrator PowerShell and restart, with virtualization enabled in the firmware:

Enable-WindowsOptionalFeature -Online -FeatureName HypervisorPlatform

Quickstart

By default, Claude, GPT and Gemini run through their CLIs, so install and log in to the one for your model: Claude Code (claude), Codex (codex) or Gemini CLI (gemini).

To call the provider's API instead, set its key and switch the org to the api backend before building the agent, e.g. ailoy.set_lang_model_provider("anthropic", "api") in Python (setLangModelProvider in Node.js, ailoy::model::set_lang_model_provider in Rust):

export ANTHROPIC_API_KEY=...   # or OPENAI_API_KEY, GEMINI_API_KEY

Amazon Bedrock, OpenRouter and vLLM are backends too; see the model providers guide.

Then build your agent with this simple API in your preferred language:

pip install ailoy-py

import asyncio

from ailoy import AgentBuilder
from virtx import ConsoleClient, Recipe, ensure_virtx


async def main() -> None:
    # Fetches virtx's server into its cache the first time; a no-op after.
    await ensure_virtx()

    console = await (
        ConsoleClient.builder()
        .image(Recipe("python:3.12-slim-trixie").step("pip install matplotlib"))
        .mount("./artifacts", "/artifacts")
        .network(True)
        .build()
    )

    agent = await (
        # For openai, use "openai/gpt-5.6-luna"
        AgentBuilder("anthropic/claude-haiku-5-5")
        .instruction("Write what you are asked for into /artifacts.")
        .system_tools()
        .console(console)
        .build()
    )

    async for output in agent.run("Create a bar chart comparing the populations of European countries and save it to /artifacts/population.png."):
        for part in output["message"]["contents"]:
            if part["type"] == "text":
                print(part["text"])

asyncio.run(main())

agent.run yields one complete message for each step of the tool loop, and agent.run_stream streams each message token by token as the model writes it.

npm install @brekkylab/ailoy @brekkylab/virtx

const { AgentBuilder } = require('@brekkylab/ailoy')
const { ConsoleClient, Recipe, ensureVirtx } = require('@brekkylab/virtx')

async function main() {
  // Fetches virtx's server into its cache the first time; a no-op after.
  await ensureVirtx()

  const console_ = await ConsoleClient.builder()
    .image(new Recipe('python:3.12-slim-trixie').step('pip install matplotlib'))
    .mount('./artifacts', '/artifacts')
    .network(true)
    .build()
  // For openai, use "openai/gpt-5.6-luna"
  const agent = await new AgentBuilder('anthropic/claude-haiku-5-5')
    .instruction('Write what you are asked for into /artifacts.')
    .systemTools()
    .console(console_)
    .build()

  try {
    for await (const { message } of agent.run('Create a bar chart comparing the populations of European countries and save it to /artifacts/population.png.')) {
      for (const part of message.contents) {
        if (part.type === 'text') console.log(part.text)
      }
    }
  } finally {
    await agent.close()
  }
}

main()

agent.run yields one complete message for each step of the tool loop, and agent.runStream streams each message token by token as the model writes it.

[dependencies]
ailoy = "0.3"
virtx = "0.1"

use ailoy::{
    agent::AgentBuilder,
    console::ConsoleClient,
    message::{Message, Part, Role},
};
use virtx::image::Recipe;
use futures::StreamExt as _;

#[tokio::main]
async fn main() -> anyhow::Result<()> {
    // Fetches virtx's server into its cache the first time; a no-op after.
    virtx::ensure_virtx().await?;

    let console = ConsoleClient::builder()
        .image(Recipe::new("python:3.12-slim-trixie").step("pip install matplotlib"))
        .mount(std::path::absolute("./artifacts")?, "/artifacts")
        .network(true)
        .build()
        .await?;

    // For openai, use "openai/gpt-5.6-luna"
    let mut agent = AgentBuilder::new("anthropic/claude-haiku-5-5")
        .instruction("Write what you are asked for into /artifacts.")
        .system_tools()
        .console(console)
        .build()
        .await?;

    let query = Message::new(Role::User).with_contents([Part::text("Create a bar chart comparing the populations of European countries and save it to /artifacts/population.png.")]);
    let mut stream = agent.run(query);
    while let Some(output) = stream.next().await {
        let message = output?.message;
        if message.role == Role::Assistant {
            for text in message.contents.iter().filter_map(Part::as_text) {
                println!("{text}");
            }
        }
    }
    Ok(())
}

agent.run yields one complete message for each step of the tool loop, and agent.run_stream streams each message token by token as the model writes it.

...Or skip the reading: point your coding agent (Claude Code, Codex, Cursor, ...) at this README and tell it what agent you want to build. This repo is its own plugin marketplace, so the agent skill installs with no clone first:

claude plugin marketplace add brekkylab/ailoy && claude plugin install ailoy@brekkylab
codex plugin marketplace add brekkylab/ailoy && codex plugin add ailoy@brekkylab

What can an agent do?

You'll need a GPU (any GPU that supports Vulkan, or Metal on Macs) for the examples that run ML models.

Example	Description
hello	One turn with no tools and no console
cad	Writes CadQuery, renders the model from four sides, looks at the renders and iterates
offshore_leaks	Analyses the ICIJ Offshore Leaks database with SQL and Python that the agent writes itself
retail_bench	Runs a supermarket simulator, one day per turn
gameplay	Plays OpenTTD (Transport Tycoon Deluxe) while you watch over VNC
sam3	Segments images and videos with SAM3 on the guest GPU (ncnn + Vulkan)
tts	Speaks text in a voice described in words, using Qwen3-TTS
laya	Answers typed decision questions with a local model on the GPU
browser-use	Waits for what you ask of it, in a window of your own browser

Each example lives in examples/<name>, with one folder per language (rust, python, node) and, where the three share files (skills, prompts, model preparation scripts), a shared folder.

Rust:

cargo run --example <name>

Python:

cd examples/<name>/python
uv run main.py

Node:

cd examples/<name>/node
npm install
npm start

How It Works

Ailoy give the agent a computer of its own.

This lets you build agents that do more than call predefined tools: they can install and use software, create their own scripts, and operate in a general-purpose computing environment—without touching the host system beyond what you explicitly expose.

To make this possible, Ailoy uses krun-based virtualization to give each agent a virtual machine of its own, with no separate VM daemon to install.

The virtualization used in Ailoy also supports the GPU, so an agent can run ML models, or even games, inside its VM.

See virtx for more details.

Building from source

git clone https://github.com/brekkylab/ailoy
cd ailoy
cargo build

For the bindings:

cd bindings/python && uv run maturin develop  # Python
cd bindings/node && npm install && npm run build  # Node.js

License

Apache-2.0. See LICENSE.md.