TALOS
One AI agent for CLI, Windows desktop and Android
Start new thread

TALOS - One AI agent for CLI, Windows desktop and Android

by•
TALOS is an AI agent ecosystem for developers across the terminal, Windows and Android. Use local GGUF models through llama.cpp, Ollama, or your own cloud provider. It combines coding tools, persistent sessions, explicit permissions, workspace checkpoints, research and on-device AI without requiring a TALOS account.

Add a comment

Replies

Best
Maker
📌
Hi Product Hunt, I built TALOS after trying a lot of AI coding environments and wanting something that wasn't tied to one provider, one model or even one interface. TALOS now has three surfaces around the same agent idea: - CLI for terminal-first development - a Windows desktop workspace - Android, including compatible GGUF models running directly on-device through llama.cpp A big part of the project is the harness around the model: explicit permissions, persistent sessions, tools, workspace checkpoints and an attempt to distinguish between “the model says it worked” and an outcome the runtime can actually verify. You can use local models, Ollama or cloud providers with your own key. I'm particularly interested in feedback on two things: 1. which surface you would actually use — CLI, Desktop or Android; 2. what you need before you trust an AI agent with a real codebase. Website: Source: Thanks for taking a look.
Maker

TALOS has just released its two biggest versions ever: desktop 0.1.25 and CLI 0.5.1.

TALOS is an open-source (AGPL) AI workspace that runs on your machine: on your projects, your files, your keys — with an agent that operates strictly within the permissions you grant.

Whats new:
🔐 Your machine, your rules. The local server now only accepts changes from the TALOS window or the TALOS mobile app, and only responds to requests addressed to its own computer. Other web pages — and other servers running on the same machine, such as a projects dev server — can no longer read or modify anything.

🤝 Agents that respect your permissions. Delegated agents work strictly within what you allow: a child agents request appears in the parent session, a yes for this session covers child agents but never extends beyond them, and agents initiating actions on their own (Coordination) is disabled by default.

⏱️ Automations from any chat. Ask any conversation to schedule work and confirm a card; the card warns you if the automation would do more than the chat it originated from.

🔒 Secrets stay secret. Search never returns the contents, names, or counts of secret files (.env, private keys, .ssh...), and reading a secret via a link prompts you first.

⚡ Faster day-to-day use: long sessions open at the latest turn, the sidebar and terminal tabs no longer freeze, and opening the reasoning box follows the models latest lines as it thinks. And it updates itself: installations on 0.1.21 and later automatically receive 0.1.25, with a signature check before anything is installed.

Try it: