How TALOS Desktop works

TALOS Desktop is a coding agent that has to show it did the work. It works in a folder of your computer, with a real terminal, a review of what it changed, automations and extensions, driven by the model you choose.

Three ideas

• Explicit authority. What the agent may do on its own is decided before it starts, and stays on screen while it works. See Permissions.
• Verified outcomes. “The model says it worked” and “TALOS observed that it worked” are two different things; only the second counts as done.
• A receipt for every action. Every operation that changes something leaves a signed receipt: what was asked, what ran, what changed.

Two parts: the window and the server

The app is a window and a local server. Everything the window does — sessions, files, Git, models, the browser — goes through the server’s local API. The server answers this computer only: it listens on the loopback address, never on the network.

• No telemetry. No usage data leaves the app.
• Settings stay here. They are kept on this computer and never synced.
• Keys stay in the system keyring, not in the window. See Models and providers.

A session is a recording

A session is a conversation with an agent that works in one folder. It is recorded as a sequence of events, not as text: you can close TALOS, reopen it and resume the session where it stopped, and the events can be replayed. What a session streams is listed in Events.

Each session shows an honest status. Those that wait for you come first; a session you stopped is neither done nor failed, it is stopped; and an old session with no recorded outcome says so instead of being counted as a success.

The first start

The first start asks four things, in this order:

1. Folder — the folder the session works in. It comes first, not last.
2. Model — which model answers.
3. Permissions — how much the agent may do without asking. Choosing Full access needs an extra confirmation.
4. Done.

The language of the window

Settings → Appearance and motion → Menu language follows your system, or is set to English or Italian. It translates menus and controls, not your messages. Menus, settings and many messages are translated; some labels are still in Italian in this release, and these pages describe them in English.

Where to go next

• Permissions — what the agent may do on its own.
• What the model reads — the preamble, project instructions and long conversations.
• Models and providers — cloud models, local models and what they cost.
• Agents and automation — sub-agents, the Board and scheduled runs.
• Where work lives — notes, tasks, the Library, memory, research and projects.