The Archive · Android · Concepts

How TALOS for Android works

An agent that lives on the phone — no account and no server of ours — with a model you choose, typed tools, and outcomes it checks.

Checked on Android 0.1.38

A newer version is out (Android 0.1.40): some details may differ.

TALOS for Android is an agent that runs on your phone. You talk to it; it answers with the model you chose, and when the answer needs it, it uses tools — your Library, your notes, the web, the phone itself — each one under a permission you set.

There is no TALOS backend. Conversations, Library files, memory and settings stay on the phone, and no account is needed to use any of it.

  • A cloud model is reached directly: what you send goes from the phone to the provider you chose, with your own key, kept in the phone’s Keystore.
  • A model on the phone answers without the network at all, and keeps working in airplane mode.
  • The web is contacted only by the operation that needs it: a search sends its words to the search service you chose; reading a page downloads that page.

Three things work with no server anywhere: models on the phone, the “hey TALOS” wake word, and everything you keep — Library, memory, notes, tasks and conversations. See Your data on the phone.

The model writes the answer; everything around it is TALOS: which tools the model is offered, whether a tool may run, what happens when a call fails or repeats, and whether an action really happened.

  • Tools are typed. Each one has a declared input, a power — read, write, or off-device — and a risk level. A tool you switch off is not sent to the model and cannot run. Every tool is listed in the tools reference, in the groups the app uses: Library, Personal, Web, Create, Models on this phone, This phone, and actions in other apps.
  • Tools are disclosed progressively. Instead of every tool’s full description at every turn, the model gets a compact catalog — or the provider’s own deferred loading, where it has one — and asks for a tool’s full schema the first time it needs it: one more round trip, and a much smaller prompt, which also helps smaller models avoid the wrong tool.
  • Every action passes a gate. Read, write and off-device are each allowed, asked or denied, and an approval is bound to the exact input it was given. See Powers and consent.

A tool saying it succeeded does not prove the world changed. Where the result can be observed, TALOS looks: for a message sent through another app, the input field emptied and the message appeared in the conversation. A result is then verified success, failed or unable to confirm — and after a timeout TALOS inspects the state instead of trying again blindly and doing the thing twice.

Android differs from phone to phone, so TALOS also keeps unsupported, denied, failed and succeeded apart: a feature this phone does not offer is not reported as an error, and an error is not reported as success.

The menu holds the chats and the app’s stations:

Station What it holds
Tasks Your to-dos, and tasks that run on their own on a schedule.
Library Files you brought, files TALOS made, and pages it read.
Models The Model Lab: providers, the catalog, and the models on this phone.
Research Deep researches and their reports.
Memory What TALOS carries into your next conversations.
Notes Notes, yours and the ones TALOS writes.
Tool Forge Tools you add from a manifest. See Tool Forge.
Code Coding sessions that run on the phone. See Start a Code session.
Doctor Checks of what works on this phone. See Check what works on this phone.
Settings Everything else, grouped as Intelligence, Connections, Interface and Privacy.

Settings also keeps an Unavailable group — Integrations, Email, Reminders and System — so that a capability the app does not have yet reads as missing, not as broken. Each one says why when you open it.

Type to search the guides.