Building one open-source AI agent across CLI, Windows and Android

I’ve been building TALOS, an open-source AI agent that I wanted to use in more than one environment.

Instead of creating a separate product for every device, the goal became:

one agent ecosystem, three interfaces: CLI, Windows Desktop and Android.

TALOS CLI

The CLI is the fastest way to use TALOS inside a development workflow.

It can work with a project, read files, run tools, keep persistent conversations and use explicit permission modes before performing sensitive actions.

Install it with:

npm install -g talos-code
talos

TALOS Desktop

The Windows application expands the same idea into a full workspace.

It combines the agent with project files, a terminal, file review, Git/GitHub workflows, persistent sessions and model configuration.

The goal is to keep the coding workflow inside one environment instead of constantly switching between an AI chat, terminal, editor and other tools.

TALOS on Android

Android is the part that originally pushed the project beyond being only a coding tool.

TALOS can run compatible GGUF models directly on the phone through llama.cpp, or connect to Ollama or cloud providers using your own API key.

A local model can therefore run without sending the conversation to a TALOS server.

Local-first by design

TALOS does not require a TALOS account and does not include TALOS telemetry.

You can choose between local models and external providers depending on the hardware and task.

I also wanted actions to be explicit: tools operate through permissions, rather than giving the agent unrestricted access by default.

Why build all three?

CLI is great when you want speed.

Desktop is better when you want visibility over files, tools and longer workflows.

Android makes the agent available away from the development machine and makes local on-device inference possible.

They solve different interaction problems, but I don’t think they need to be completely separate AI products.

That is the experiment behind TALOS.

The project is open source under AGPL-3.0.

Website: https://www.talos-code.com/

Source: https://github.com/Ninozzz95/talos

I’d especially like feedback from people building coding agents or running local LLMs:

What would you need to trust an agent enough to use it on a real project?